The Idealization of Causation in Mechanistic Explanation Alan C. Love∗ and Marco J. Nathan†‡ POBAM Draft, March 2014

Abstract Causal relations among components and activities are intentionally misrepresented in the mechanistic explanations found routinely in the life sciences. Since these causal relations are the source of the explanatory power ascribed to descriptions of mechanisms, and advocates of mechanistic explanations explicitly recognize the importance of an accurate representation of actual causal relations, the reliance on these idealizations in explanatory practice conflicts with the stated rationale for mechanistic explanations. We argue that these idealizations signal an overlooked feature of reasoning in molecular and cell biology—mechanistic explanations do not occur in isolation—and suggest that explanatory practices within the mechanistic tradition share commonalities with the model-based approach to science prevalent in population biology.

∗ Department of Philosophy, Minnesota Center for Philosophy of Science, University of Minnesota † Department of Philosophy, University of Denver. ‡ Both authors contributed equally to this work

1

1

More Thoughts About Mechanisms

The concepts of mechanism and mechanistic explanation have been the object of much recent attention in philosophy of science (see, for example, Darden 2006; Craver 2007; Fagan 2012). As it often happens, this increased scrutiny has had a somewhat polarizing effect. On the one hand, enthusiastic supporters suggest that thinking about mechanisms sheds light on many central philosophical debates in the philosophy of science, such as causation, explanation, reduction, and emergence (Glennan 1996; Machamer et al. 2000; Bechtel 2011). For instance, it is sometimes claimed that the ‘open-endedness’ of mechanistic explanations—which are not limited to linguistic representations or logical inferences but may involve diagrams and simulations—constitutes a substantial advantage over deductive-nomological explanations (Bechtel and Abrahamsen 2005). On the other hand, critics of the so-called ‘new mechanistic philosophy’ have raised some skeptical concerns, arguing that these concepts are insufficiently characterized or suffer from distinctive problems. For example, systems biology and neuroscience provide explanations that allegedly violate two standard features of mechanistic explanation: localization and decomposition (Silberstein and Chemero 2013). Regardless of differences in where people stand in this debate, one can hardly deny that mechanistic approaches have entered the mainstream of philosophy of science, in no small part due to a constant appeal to mechanisms in scientific literature. The aim of this article is to focus attention on a neglected feature of mechanistic explanations, understood as adequate descriptions of mechanisms: the idealization of causation. One of the purported advantages of concentrating on mechanisms, especially in many areas of life science, is the ability to flesh out a notion of causal explanation that corresponds to the actual practices of working scientists. In particular an account of mechanistic explanation should clarify why the common appeal to causal relations among entities and activities organized in a mechanism is explanatory.1 We argue that the intentional mis1 There

is a longstanding debate about whether all explanation is causal. Here we only

2

representation of these causal relations—the source of the explanatory power in the mechanistic description—generates a significant problem for the mechanistic framework. We begin by reviewing the typical representation of causal relations in molecular explanations (§2) and show how this widespread practice has puzzling implications for standard mechanistic accounts that have thus far been overlooked (§3). Next, we consider several plausible responses to this situation that ultimately fail (§4). We conclude by proposing an alternative solution that exposes a commonality between mechanistic explanations in molecular biology and modeling in population biology: the employment of multiple, idealized models to account for the behavior and properties of living systems (§5).

2

Causal Relations in Mechanistic Explanation

Providing a concise—yet informative—definition of ‘mechanistic explanation’ is no easy task. Part of the problem has to do with substantial disagreements regarding what counts as a mechanism, as different proponents of a ‘new mechanistic philosophy’ adopt distinct and incompatible views of the relevant notions (Woodward 2013). Whereas Craver (2008) embraces a more restricted conception of mechanism, which emphasizes the importance of providing as much detail as possible in its description, Bechtel (2011) prefers a more ecumenical interpretation, according to which various different kinds of explanations count as ‘mechanistic.’ In an attempt to keep our analysis as general as possible and to avoid getting entangled in technical disputes, we treat mechanistic explanation as the claim that many areas of science explain by decomposing systems into their constituent parts, localizing their characteristic activities, and articulating how these are organized to produce a particular effect. Thus, instead of providing a systematic account of mechanism or explicating the basic concepts of a mechanistic approach, such as ‘organization’ or ‘activity,’ our emphasis will be on the core conception: mechanistic explanations illustrate and display the assume that causal explanations constitute an important subclass of explanations in science that are especially salient in discussions of mechanisms.

3

Figure 1: A pictorial depiction of the mechanism for LTD inhibition. generation of specific phenomena by describing the organization of a system’s constituent components and activities. This emphasis is intended to capture a common ground among various new mechanists regarding the nature of mechanistic explanation. Let us begin by considering how causal relations are represented in mechanistic models. Mechanistic explanations in molecular and cell biology typically involve both a verbal description and a pictorial depiction of the relationships among constituents. To illustrate, consider a paradigmatic example—the mechanism for LTD inhibition (Figure 1).2 The inhibition of a lasting reduction in synaptic transmission—Long Term Depression (LTD)—is accomplished through an intercellular mechanism that is initiated by the lasting enhancement of synaptic transmission—Long Term Potentiation (LTP). The explanandum phenomenon is produced by interactions between component molecules that activate or repress other molecules. More specifically, in N-methyl-D-aspartate receptor (NMDAR) dependent LTD, NMDAR activation induces a signaling cascade that activates Phosphoinositide 3-kinase (PI3K) which, in turn, acti2 The image is borrowed from http://www.bris.ac.uk/synaptic/research/projects/mechanisms. The descriptive details of the mechanistic explanation are based on Peineau et al. (2007) and Li et al. (2010).

4

vates Protein kinase B (Akt). Akt deactivates Glycogen synthase 3 isoform β (GSK-3β) through the phosphorylation of one of its serines (an amino acid residue in the protein), which prevents the internalization of the α-amino-3hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA) receptor. If Caspase-3 is present, then it cleaves Akt (a form of deactivation). This leaves GSK-3β unphosphorylated, which then facilitates the internalization of the AMPA receptor and subsequent establishment of LTD. The specific details of this complex molecular process do not matter here. The important point is that this type of representation of component interactions and activities in mechanisms is ubiquitous in biological explanations. Causal relations are typically represented by arrows, sometimes with ‘+’ signs denoting initiation or activation and ‘−’ signs denoting inhibition or repression. The depicted causal relations are simplified and reflect only a small subset of those occurring in the cellular context. The arrows in Figure 1 do not distinguish among different kinds of biochemical interactions and ignore various background conditions, such as features of the cytological environment or the exact duration of the processes. These arrows simply stand in for causal relations, regardless of how they are instantiated. In this sense, mechanistic explanations involve abstraction—the intentional omission of detail—evident in the practices of displaying the described components as distinguishable geometric shapes or the exclusion of known components (such as Ca2+ /calmodulin-dependent protein kinase II activation that occurs subsequent to NMDAR activation in parallel with PI3K). Abstraction must be carefully distinguished from idealization—the deliberate misrepresentation of detail in a model. Paraphrasing Godfrey-Smith (2009), abstract descriptions ‘leave out a lot,’ whereas idealized descriptions ‘fictionalize in the service of simplification.’ A vector representation of forces in physics constitutes an abstraction from a variety of real processes; the notion of an ideal gas intentionally misrepresents certain properties of gases, such as the shape and the interactions between molecules. In the LTD inhibition mechanism presented above, the arrows represent a one-one relationship between molecular compo5

nents and their associated activities in an ordered sequence (for example, a unit of PI3K activates a unit of Akt). Note, however, that the diagram does not simply let types stand in for tokens (a common form of abstraction) but idealizes the causal relationship. The downstream consequences of LTD inhibition do not follow automatically from a molecule of PI3K activating a molecule of Akt, as depicted in the model: particular concentrations of PI3K are required before activation or repression occurs. Although the significance of both abstraction and idealization in model building and explanatory practices is well known (Weisberg 2013), the literature on mechanistic explanation has stressed the former (Levy and Bechtel 2013). This is striking, especially given that these idealizations are localized to the exact place in the mechanism description where the explanatory force obtains—the causal relations.

3

Causal Relations Idealized

That mechanistic explanations are both abstract and idealized is unsurprising. The cell and its myriad constituents compose an extremely sophisticated and complex apparatus that includes a diversity of components and a multitude of interactions. A realistic representation of this plethora of entities and relations— assuming that such as ‘complete’ description is even feasible—would make the description impractical and the explanation unilluminating. Abstraction and idealization are necessary aspects of explanatory models: features that do not play a central role in the explanation can (and should) be abstracted away or distorted in order to make the representation more perspicuous. Mechanistic models are no exception; they should contain all and only the core explanatory components and activities of the mechanism. Although determining exactly which elements ought to be included or depicted accurately in the explanation constitutes a substantial question (Strevens 2008), the basic criterion of inclusion is simple: it involves deciding whether an accurate or explicit description of the element contributes to the explanation. An important corollary of this principle is that we would not expect the features that played a central role in 6

the explanation to be abstracted away or idealized in a mechanistic description. Yet, this is exactly what occurs in mechanistic explanations in molecular biology: the causal relations represented by arrows in the depiction of mechanisms that account for the explanandum phenomenon are deliberately misrepresented. To illustrate the nature of the problem, consider gene expression, one of the most important and better-studied cytological phenomena, where a gene (that is, a portion of DNA) is transcribed into RNA. This RNA is then translated into a protein product that can interact with distinct biochemical moieties or is active in its own right within specific cellular processes, as in the case of enzymes or regulatory molecules. Descriptions of gene expression are good examples of the paradigm of mechanistic explanation presented by Robins and Craver (2009). First, the explanandum phenomenon is well understood and formulated precisely: the goal of the explanation is to account for the biochemical process responsible for transcribing a given sequence of nucleotides into RNA within the context of the cell. Second, both the component entities and their activities have been thoroughly investigated and can be described in detail. Biologists have successfully identified most of the specific molecular entities that are necessary for gene expression, such as operon sites and transcription factors, as well as the relevant background conditions, such as the presence of specific enzymes required for a particular gene to be expressed (activated) at a given time. Finally, the structure of the system is well-defined: both the spatial and temporal organization that produces the explanandum (gene expression) can be specified and modeled with impressive accuracy, including details regarding the location, size, shape, and orientation of the relevant components in different processes, as well as their order, rate, and duration. Despite the impressive strides achieved over the last few decades, our current knowledge of how gene expression works remains incomplete and contemporary biology is constantly uncovering additional details of the mechanism. As a result, current accounts of gene expression are naturally viewed as mechanism schemas with placeholders that indicate where further details need to be supplied (Darden 2006). These ‘incomplete’ models of mechanisms fall short of an 7

Figure 2: A diagrammatic representation of gene expression. A transcription factor molecule binds to the DNA at its binding site, and thereby regulates the production of a protein from a gene. ideally complete mechanistic description, but are nonetheless explanatory: the available mechanism descriptions facilitate deeper predictions and can be verified and controlled through surgical experimental manipulations. Despite these inevitable empirical shortcomings, gene expression is one of the most studied and best-known mechanisms in molecular biology. If gene expression is not understood and explained mechanistically in a robust sense, then it is hard to see what else would fit the bill. How is causation treated in these mechanistic models of gene expression? A standard diagrammatic representation (Figure 2) shares both the abstractions and the idealizations exemplified in the mechanistic description for LTD inhibition (Figure 1). In this model, causal relations are indicated by proximity: the binding of a transcription factor on the site upstream of transcription initiation regulates the gene by triggering the transcription of DNA into RNA and, subsequently, the translation of RNA into protein. These kinds of diagrammatic representations are common in textbooks. However, in more advanced discussions, we find increasingly detailed representations of eukaryotic gene expression and more precise narrative descriptions of the mechanism (Figure 3; Ptashne 8

Figure 3: A more detailed diagrammatic representation of gene expression (Levine and Tijan 2003). and Gann 2002). This more specific account of the apparatus for the regulation of eukaryotic gene expression exposes a variety of abstractions that were present in Figure 2. For instance, Figure 3 emphasizes that transcription factors operate in conjunction when binding to the upstream promoter region (represented by blue-colored ovals, such as TFIID), and also require the operation of cofactors (represented by peach-colored ovals: CRSP, ARC). These crucial components were omitted in Figure 2. Importantly, Figure 3 is still far from a complete description: other, even more specific, depictions could fill out the omitted details (e.g., by specifying intermediary steps or further necessary components, such as enzymes that catalyze biochemical reactions). Yet—and this is the crucial point—the lack of further detail does not undermine the explanatory force of these mechanistic characterizations: the goal of these (deliberately) oversimplified diagrams is to identify the core features—and these features only—of the mechanism of gene expression. Differences in the particular level of incorporated detail depend on the explanatory goals of researchers in concrete investigative contexts. The features of mechanistic models presented so far should be neither controversial nor problematic. Trouble begins to arise when we combine the above discussion with the assumption, commonplace among many philosophers of sci-

9

ence, that an adequate causal explanation (whether or not it is ‘mechanistic’) should include all and only the difference makers that ensure the production of the explanandum phenomenon (Waters 2007). This generates a tension because the way in which the various components of the mechanism purportedly make a difference as to whether or not the gene is transcribed are misrepresented in the diagrams of gene expression. Consider the green-colored oval labeled ‘transcription factor’ in Figure 2, which stands for the binding of a transcription factor to the binding site. This single feature of the diagrammatic representation contains three important misrepresentations. The first misrepresentation results from the abstraction that the site binds a single molecule; as revealed in Figure 3, what the site binds is rather a complex of molecules. While this might seem like a pedantic observation, especially given that the abstraction was already acknowledged, it is a key point for understanding the mechanism of gene expression. In most circumstances, individual molecules do not act as difference-makers, but complexes of multiple molecular subunits do (a point we elaborate further below). Thus, this representation in the mechanism description does not simply ‘leave things out’ but ‘fictionalizes in the service of simplification.’ The second misrepresentation in the diagram is that gene expression is depicted as being triggered by a single transcription factor or, more precisely, a single complex of transcription factors—let us call this molecular unit (or its associated complex) ‘p1 .’ While p1 unquestionably plays a role in the process, it is not a difference maker by itself; its presence or absence makes virtually no difference to the outcome. This is because if p1 was not there, another molecule of the same type (p2 , p3 , . . . , p546 , . . .) would take its place. (Parallel reasoning applies to the multi-unit complex, though different tokens of different types increase the combinatorial permutations.) Some readers might protest that the diagram is supposed to capture sufficient—not necessary—conditions for transcribing the gene. Furthermore, the transcription factor oval is construed as a type, not as a token; it is meant to represent any molecular complex that attaches to the binding site. While there 10

are many alternative ways in which the necessary conditions could be fulfilled, the diagram does what it is intended to do, and does so well: it captures (in an oversimplified form) the sufficient conditions under which the gene in question is actually expressed. The problem with this response is that it overlooks an important feature of the binding process: transcription factors (or associated complexes) do not remain attached indefinitely to the operon site once they bind; rather, they are constantly dislodged and their place is taken by other nearby molecule tokens of the same type. The regulation of gene expression thus operates through an extended temporal interval because multiple tokens of a transcription factor need to bind to the switch consecutively, in order for the gene to be expressed. This series of bindings cannot be represented—as Figure 2 purports to do—as an individual binding event; what is required is rather a matrix representing a variety of binding events (Nathan ming). Individual molecules or single molecular complexes are not difference makers: they represent neither necessary nor sufficient conditions for gene transcription. One might grudgingly accept these idealizations, despite the fact that they inaccurately depict how the mechanism works, but there is a third misrepresentation that amplifies the difficulty. Based on our discussion, the diagram in Figure 2 represents some components that do not make a difference to the outcome (i.e., the expression of the gene). But then what are the difference makers underlying the process? What actually makes a difference as to whether or not the gene is transcribed is the concentration of transcription factor present in the entire system (Nathan 2012).3 The problem is not simply that the diagram introduces component entities that do not make a difference—the diagram fails to incorporate some features of the system that do play an actual differencemaking role. In order to avoid this difficulty, one might reply that the transcription factor oval is shorthand for the entire concentration of transcription factor and, therefore, implicitly represents the pattern of actual bindings. This response 3 More precisely, the difference maker is a relative concentration—the concentration of transcription factor relative to the concentration of repressor that would inhibit the transcription of the gene were it to bind to the operon site.

11

is inappropriate for two reasons. First, the claim that the transcription factor oval stands in for an entire concentration of molecules is inconsistent with the description associated with the mechanism diagram: “a transcription factor molecule binds to DNA at is binding site, and thereby regulates the production of protein from a gene” (emphasis added). This verbal representation of the gene expression mechanism does not invoke concentrations. Second, the pictorial representation of a transcription factor type standing in for transcription factor tokens does not capture the notion of a concentration. Focusing on the single molecule situation, for the sake of simplicity, suppose that a concentration of one hundred molecules is necessary to make a difference and activate transcription. The diagram does not capture this information, either explicitly or implicitly, because the single abstract geometrical shape depiction is perfectly consistent with concentrations that do not make a difference (e.g., twenty-five molecules). Biologists are aware of the significance of concentrations for difference-making causal relations in cellular contexts (Ptashne 2004); the mechanism description deliberately misrepresents them. Concentrations of transcription factors are nowhere mentioned in these diagrams or textbook explanations of gene expression because the difference-making causal relations in mechanistic diagrams are idealizations: they intentionally ignore known variations in properties and exclude particular values of variables that account for why the explanandum phenomenon occurs. This situation becomes even more puzzling when we note that it clashes with a criterion of adequacy for mechanistic explanations that is often stated and endorsed explicitly: How-possibly models are often heuristically useful in constructing and exploring the space of possible mechanisms, but they are not adequate explanations. How-actually models, in contrast, describe real components, activities, and organizational features of the mechanism that in fact produces the phenomenon. They show how a mechanism works, not merely how it might work (Craver 2007, 112).

The organized relationships between components and activities that demon12

strate how a mechanism actually produces a particular phenomenon are idealized— they do not show how the mechanism actually works. Particular concentrations of molecular species are required before activation or repression occurs and, therefore, the arrow drawn between two abstract components (that is, geometrically distinguishable shapes) is a deliberate misrepresentation. The dilemma should now be apparent. A widespread idealization practice commonly utilized in the description of mechanisms—idealizing causal relations— conflicts with the explicit goal of accurately representing actual causal relations in mechanistic explanations. Biologists engage in a series of abstractions and idealizations that result in the deliberate misrepresentation of the productive continuity between difference-makers that is the hallmark of mechanistic explanation. Purported difference makers are represented in such a way that they are not difference makers according to what is known about the mechanism. The idealization of causal relations—the intentional misrepresentation of how the mechanism produces the phenomenon—means that these models do not show how the mechanism actually works. Mechanistic explanations thus appear to fail according to their own criteria.

4

Objections and Replies

There are various strategies that proponents of mechanistic explanation can adopt to address or dismiss the puzzle of idealizing causal relations. As several authors have recognized, the new mechanistic philosophy tends to be divided on one fundamental point. Whereas one set of authors emphasizes the importance of providing mechanistic models that are as complete and specific as possible (Machamer et al. 2000; Darden 2006; Craver 2007), other have recognized— more or less explicitly—the significance of abstracting away from unnecessary details (Bechtel 2011). Depending where one falls on this spectrum, there are (at least) two different strategies that a mechanist could employ to address this problem. Philosophers who value the completeness and richness of detail in mecha13

nistic explanations are likely to respond by drawing a distinction between ideally complete mechanistic descriptions and mechanisms schemas (or mechanism sketches). As noted, pictorial representations such as Figures 1 and 2 fall short of complete descriptions: there are various processes and entities contributing to gene expression that do not figure in the model. In response to this situation, some mechanists could say that such representations are merely sketches of mechanisms, which need to be filled out further in order to be fully explanatory. On this view, the idealization of causation only appears to be a puzzle because of our current lack of knowledge. Once all the details have been figured out, we will be able to accurately specify all of the constituents and the various processes connecting them, and no difference makers will be idealized. This reply has the merit of emphasizing an important aspect of scientific explanation: it is still possible to explain phenomena when important knowledge is lacking by using ‘filler-terms’ or black-boxes that function as placeholders until more precise details of the structure of a mechanism are discovered.4 At the same time, this response is ultimately unconvincing because the gradual elimination of idealized diagrams is rarely—if ever—witnessed in scientific practice. Indeed, even when the relevant details are known, as in the case of gene expression, researchers do not bother to replace idealized causal relations with more realistic representations that provide a more accurate depiction of the causal relations responsible for the explanandum phenomenon. The authors of Figure 2 know perfectly well how difference making operates from the transcription of a sequence of nucleotides to the production and folding of a protein. Still, they deliberately choose to abstract away from some details and idealize others to keep the diagram simple and perspicuous. Furthermore, even when additional details are provided, as in Figure 3, the relevant idealizations remain unacknowledged and uncorrected (in part because they require different representational modes). While mechanists committed to the explanatory virtue of completeness 4 The history of science is replete with explanatory accounts that initially set important phenomena aside, such as Darwin’s black-boxing of the mechanisms of ontogeny or the attempts of early psychologists to explain mental processes while ignoring the underlying neural processes.

14

might view this as a flaw in the explanatory practice of biology, we argue that it signals an important—albeit overlooked—feature of reasoning in molecular and cell biology: mechanistic explanations do not occur in isolation (see below, §5). Not all mechanistic philosophers are committed to the explanatory virtue of completeness. Another group of mechanists emphasizes the importance of abstractions in the representation of mechanisms, and would likely adopt a difference response to the problem of idealizing causal relations. For example, Levy and Bechtel (2013) argue that one needs to abstract away from the structural specifics of a mechanism and represent it in a skeletal, coarse-grained matter in order to understand its organization: “[I]t is often the connectivity, treated abstractly that explains why a mechanism exhibits the particular behavior it does” (245). This attention to a key role for abstraction in mechanistic explanation is welcome but it fails to resolve the difficulty. Although philosophers of science tend to address abstraction and idealization in similar ways, these two features pose different and independent issues for mechanistic explanations. The widespread use of ‘irreducible’ abstractions must be reconciled with the claim that ideal mechanistic explanations should be as accurate and realistic as possible; this can be accomplished by acknowledging that not all causal features of the mechanism need to figure in the explanatory representation. Abstractions are compatible with the goal of describing how a mechanism actually works because they play the role of making the model and, therefore, the explanation, more perspicuous. In contrast, idealizations cannot be accommodated so easily within the mechanistic framework, since they overtly violate the ‘actuality’ requirement. Idealizations introduce deliberate falsities or misrepresentations into mechanistic explanations, which involves a direct clash with the claim that mechanistic representations should represent how systems (or their subcomponents) actually work. A mechanist could respond by relaxing the actuality criterion and denying that realistic descriptions of mechanisms are required to achieve mechanistic explanations. But once we give up on realism, it becomes unclear what exactly is doing the explanatory work in mechanistic descriptions.

15

5

Multiple Modeling in Molecular Biology

In the first part of this paper, we argued that accounts of mechanistic explanation face a problem in accommodating the deliberate misrepresentation of causal relations among the components and activities of a mechanism that play a central difference-making role in producing the explanandum phenomenon. Two distinct responses to this problem were reviewed in the previous section: (i) idealizations will eventually be removed as our knowledge of the mechanism increases and idealized descriptions will be progressively replaced by more accurate and realistic depictions; and, (ii) the normative goal of providing the most realistic explanations that adequately represent causal relations in a mechanism can be abandoned. The first response does not fit with the practice of scientists who do not progressively eliminate idealizations, even in cases where more accurate models are readily available. The second response makes it difficult to understand how a mechanism description operates as a causal explanation. How then should we interpret the idealization of causation in mechanism descriptions? One promising approach is to recognize that mechanistic explanations do not occur in isolation. This strategy has been neglected, we surmise, for two reasons. First, the solution requires distinguishing different—but equally important—forms of idealization that are often lumped together. Second, the problem we have explicated weakens ontological commitments to mechanisms that several philosophers regard as important. Weisberg (2013, Ch. 6) distinguishes between three different kinds of idealization in modeling. The first kind is what he dubs Galilean idealization: the practice of introducing distortions for the sake of simplifying theories. Galilean idealization is especially important in areas of science, such as computational chemistry, that deal with complex systems and require simplifying assumptions in order to render their theories computationally tractable. Despite their prominence in various scientific fields, Galilean idealizations are not germane to the pictorial depictions of mechanisms. This is because the practice is largely pragmatic—theorists idealize for reasons of computational tractability that are

16

a function of our cognitive (in)abilities—and non-permanent since the idealizations take place with the expectation of future de-idealization and more accurate representation. While the expectation of future de-idealization fits in well with the first mechanistic strategy of interpreting idealizations as mechanism schemas, the pragmatic dimension of Galilean idealization is hard to reconcile with the realistic aim underlying most accounts of mechanistic explanation. For these reasons, most discussions of mechanistic explanation more or less explicitly assume that mechanistic descriptions involve a second, independent kind of idealization, which Weisberg calls minimalist idealization. Minimalist idealizations introduce only those core causal factors that make a difference in producing a phenomenon or its essential features, a desideratum that has been refined and endorsed by recent accounts of causal explanation (Strevens 2008). It is important to note that the product of Galilean and minimalist idealizations need not be different; following these two strategies can lead to the production of identical models with idealizations of precisely the same features. But there would still be an important difference in the way that these models are justified (Weisberg 2013, 105ff). Specifically, whereas a Galilean idealizer will claim that the model introduces the variables it does because of pragmatic usefulness, a minimalist would claim that the main rationale for adopting the model is that it captures the causally relevant factors. Mechanists tend to prefer a minimalist approach to idealization over a Galilean one because the former avoids the pragmatic flavor explicit in the latter. Furthermore, one should not expect minimalist idealizations to be gradually eliminated as science progresses, making the view more consistent with what we witness in actual scientific practice. At the same time, minimalist idealization fails to adequately address the problem of idealizing causal relations. This is because it is hard to reconcile the general goal of minimalist idealization—to include all and only those factors that make a difference to the production of the explanandum phenomenon—with the idealization of precisely those causal relations that account for the explanandum. Neither Galilean nor minimalist idealizations solves the puzzle of idealized representations of causation in mechanism descriptions. 17

The third type of idealization introduced by Weisberg is multiple-model idealization, which involves building multiple related but incompatible models that capture distinct aspects of the causal structure generating complex phenomena. This practice has many parallels with the second mechanistic strategy of relaxing the actuality criterion for mechanism descriptions. The key difference is minimalist idealization does not aim to produce a single best model. Instead, multiple models with different idealizations and abstractions are constructed so that an understanding of how the explanandum phenomenon is produced results from comparing and contrasting the different models. Although new knowledge can be added to each model individually, there is no expectation that the idealizations will be progressively removed or that the need for multiple models will fade over time. The recognition that mechanistic explanations cannot be subsumed under a single overarching model helps to secure a solution to the problem of idealized causal relations in mechanistic explanation. A multiple-model idealization strategy addresses each of the three misrepresentations identified in mechanistic descriptions. First, it accounts for why it is often possible to offer more accurate models by relaxing abstractions that engender misrepresenting causal relations even though more coarse-grained models are not eliminated and replaced. Different representations of gene expression (Figures 2 and 3) provide alternative depictions that correct for the fact that the site actually binds a complex of molecules, as opposed to a single molecule or a single molecular unit. But Figures 2 and 3 are not competing descriptions; they coexist. Each model represents different idealizations and abstractions with respect to gene expression. Additionally, these depictions are not meant to be exhaustive: additional models are required to account for how these difference-making relations obtain. In this respect, multiple-model mechanistic explanations in molecular and cell biology are closely associated with theorizing about population phenomena in ecology and evolution, which has been termed ‘model-based science’ (GodfreySmith 2006). In both cases, it is a mistake to assume that more accurate models should always replace less accurate models. Descriptive accuracy is a virtue of 18

mechanistic explanation, but one that can often only be obtained at the cost of sacrificing another virtue: generality. The second insight provided by the multiple-modeling approach to idealization involves showing how it is possible to offer models that are dynamic with respect to transition events in mechanistic descriptions. Introducing a stochastic binding model to represent this characteristic feature of cellular environments can capture that genes are activated by concentrations of transcription factors, as opposed to individual molecules. Such a model would depict the process more accurately, but it cannot replace the original ‘deterministic’ one because it would entail a significant loss of perspicuity and generality. This accounts for why few, if any, introductory textbooks present stochastic binding models. The reason is not, as it is often assumed, that these models are too complex for beginning scholars. It is that they are too specific: the simpler, deterministic models are more general and are therefore applicable to a greater variety of phenomena. However, stochastic binding models can (and should) be used to complement deterministic ones in contexts where it becomes important to provide more accurate descriptions of the binding process. Third, and finally, the multiple-modeling approach allows us to offer models that explicitly treat idealized difference-making relations as separate mechanisms, which usually requires that we utilize mathematical equations to model concentration effects. The important lesson is that the deterministic model of gene transcription (where a single molecular units binds to the operon site to trigger gene expression) and the stochastic model that factors in concentrationeffects and other potential difference-makers, are not alternative descriptions of the same mechanism. Rather, they are different mechanisms that capture and explain different features of a process. Once we treat such depictions as different mechanisms, it becomes clear why asking which representation is better or more accurate is an ill-posed question; decisions about inclusion and degree of accuracy in a representation will depend on the explanandum in view. In sum, the misrepresentation of difference-making causal relations in mechanism descriptions is a pervasive feature of explanatory practice in molecular 19

and cell biology that is in direct conflict with the expectation that an accurate representation of actual causal relations is the source of their explanatory power. We have shown that this problem can be addressed by interpreting the idealizations in mechanistic descriptions as part of a multiple-modeling strategy. This means that the goal of mechanistic explanation is not a single model that contains all of the relevant causal details, but rather a series of multiple, complementary, mechanistic diagrams and descriptions comprised of different idealizations, similar to what is observed in population biology. Our claim is not that multiple-modeling idealization is the only significant form of idealization, or even the most important one. Weisberg is correct that all three forms of idealization (and possibly more) play significant and independent roles in scientific practice. Our claim is that once we understand the idealizations present in mechanistic models along the lines suggested by the multiple-modeling interpretation, the puzzle of idealized causal relations can be resolved.

6

Concluding Remarks

In closing, we offer two general remarks, one regarding the analogies between molecular and population biology, and another regarding the metaphysical implications of our discussion. Abstraction and idealization are essential and irreducible features of scientific representation. This is because describing and explaining are different scientific goals that involve a necessary tradeoff between explanatory power and descriptive accuracy (Cartwright 1983). Descriptive accuracy encourages focusing on individual token-events, whereas explanation presupposes the classification of discrete event-tokens under a more general type. Whenever distinct tokenevents are subsumed under the same type, this requires abstracting away or idealizing the specific differences that distinguish the two events. The broader and more diverse the class of events subsumed under the same type, the more radical the abstractions and idealizations required to classify them together. This is precisely the situation biologists confront when dissecting how molecu20

lar mechanisms operate. One can generate an explanation of the specific features that have been misrepresented in the original mechanism description, but doing so requires supplying a distinct and independent model that contains different idealizations and abstractions, as well as a more accurate and realistic depiction of the causal relations misrepresented by arrows in the first mechanistic model. Together, the models increase the overall explanatory power. Arrows connecting two molecular types to account for an explanandum phenomenon function as placeholders that can be refined and replaced with a new distinct mechanistic model, which depicts complex causal relations more accurately and realistically (such as concentrations of transcription factors or other populations of molecular entities). However, the new model does not replace the previous one because in order to increase descriptive accuracy it will necessarily lose some of its generality. As noted above, one major reason why mechanist philosophers tend to assume—more or less explicitly—a minimalist approach to idealization is to preserve the ontological import of mechanistic explanations: “explanations [mechanisms] are objective features of the world” (Craver 2007, 27). Minimalist idealizations can be made perfectly consistent with this ‘objectivist’ approach to mechanisms because idealizations fulfill a role analogous to abstraction; they take out or simplify irrelevant details. In contrast, the multiple-model approach suggested here is harder to reconcile with this objectivity since inconsistent models cannot both be true simultaneously. Therefore, an awareness of multiplemodel idealizations of causal relations in mechanist explanation should temper any ontological implications drawn from mechanistic models. This constitutes a meaningful lesson for philosophy of science: distilling metaphysical implications from science requires carefully attention to explanatory practice. This broader methodological lesson requires further development but, at a minimum, our analysis of a widespread but neglected feature of mechanistic explanation—the idealization of causation—provides a novel point of departure for comprehending the complex nature of ubiquitous appeals to mechanisms in the explanations offered by molecular and cellular biologists. 21

References Bechtel, W. (2011). Mechanism and biological explanation. Philosophy of Science 78, 533–57. Cartwright, N. (1983). How the Laws of Physics Lie. Oxford: Clarendon. Craver, C. F. (2007). Explaining the Brain: Mechanisms and the Mosaic Unity of Neuroscience. Oxford: Clarendon Press. Craver, C. F. (2008). Physical law and mechanistic explanation in the Hodgkin and Huxley model of the action potential. Philosophy of Science 75, 1022–33. Darden, L. (2006). Reasoning in Biological Discoveries: Essays on Mechanisms, Interfield Relations, and Anomaly Resolution. Cambridge, UK: Cambridge University Press. Fagan, M. B. (2012). The joint account of mechanistic explanation. Philosophy of Science 79, 448–72. Glennan, S. (1996). Mechanisms and the nature of causation. Erkenntnis 44 (1), 49–71. Godfrey-Smith, P. (2006). The Strategy of Model-Based Science. Biology and Philosophy 21, 725–40. Godfrey-Smith, P. (2009). Darwinian Populations and Natural Selection. Oxford, UK: Oxford University Press. Levine, M. and R. Tijan (2003). Transcription regulation and animal diversity. Nature 424, 147–51. Levy, A. and W. Bechtel (2013). Abstraction and the organization of mechanisms. Philosophy of Science 80 (2), 241–61. Li, Z., J. Jo, J. Jia, and et al. (2010). Caspase-3 activation via mitochondria is required for long-term depression and AMPA receptor internalization. Cell 141, 859–71. Machamer, P. K., L. Darden, and C. F. Craver (2000). Thinking about Mechanisms. Philosophy of Science 67, 1–15. Nathan, M. J. (2012). Causation by concentration. British Journal for the Philosophy of Science (doi:10.1093/bjps/axr056), published online. Nathan, M. J. (forthcoming). Redundant causality and explanatory robustness. Minnesota Studies in the Philosophy of Science. Peineau, S., C. Taghibiglou, C. Bradley, and et al. (2007). LTP inhibits LTD in the hippocampus via the regulation of GSK3. Neuron 53, 703–17.

22

Ptashne, M. (2004). A Genetic Switch. Phage λ Revised (3rd ed.). Cold Spring Harbor, NY: Cold Spring Harbor Laboratory Press. Ptashne, M. and A. Gann (2002). Genes and Signals. Cold Spring Harbor, NY: Cold Spring Harbor Laboratory Press. Robins, S. K. and C. F. Craver (2009). Biological clocks: Explaining with models of mechanisms. In J. Bickle (Ed.), The Oxford Handbook of Philosophy and Neuroscience, pp. 41–67. New York: Oxford University Press. Silberstein, M. and A. Chemero (2013). Constraints on localization and decomposition as explanatory strategies in the biological sciences. Philosophy of Science 80 (5), 958–70. Strevens, M. (2008). Depth. An Account of Scientific Explanation. Harvard University Press. Waters, C. K. (2007). Causes that Make a Difference. The Journal of Philosophy 104 (11), 551–79. Weisberg, M. (2013). Simulation and Similarity: Using Models to Understand the World. New York: Oxford University Press. Woodward, J. (2013). Mechanistic explanation: Its scope and limits. Proceedings of the Aristotelian Society 87, 39–65.

23