2.1. Explanatory Unification

The defense of explanatory unification as a virtue of selectionist evolutionary theory has roots in Darwin's use of Whewell's notion of consilience. Darwin (Reference Darwin1903) wrote: “I have always looked at the doctrine of natural selection as an hypothesis, which if it explained several large classes of facts, would deserve to be ranked as a theory deserving acceptance” (1:139–140). This virtue of selectionist evolutionary theory continues to de defended (see, e.g., Ruse Reference Ruse1979; Kitcher Reference Kitcher1981; for opposing arguments, see Lloyd Reference Lloyd1983; Hodge Reference Hodge and Ruse1989). The theory productively and reliably brings together disparate types of evidence—hybridization, developmental, morphological, biogeographical, and paleontological phenomena are placed under a single theoretical frame. Selectionist evolutionary theory conjoins multiple kinds of old evidence.

Consilience is a nontrivial achievement. Successful unification provides confirmation “boosts” for each piece of the theoretical structure (e.g., Friedman Reference Friedman1974, Reference Friedman and Healey1981; Kitcher Reference Kitcher1981, Reference Kitcher, Salmon and Kitcher1989). This is because confirmation is acquired indirectly: the empirical adequacy of one part of the theory strengthens a distinct, but connected, part.Footnote ¹ Moreover, our confidence in the empirical adequacy of the whole theory also increases as a result of unification.Footnote ² Lynn Margulis (Reference Margulis1975) and E. O. Wilson (Reference Wilson1998) speak to the importance of unification in evolutionary theory.

2.2. Model Fitting

The defense of model fitting as a theoretical virtue of evolutionary theory stems mainly from twentieth-century population genetics. Common statistical procedures for model fitting include (1) regression analysis and (2) Neyman-Pearson techniques for hypothesis testing (e.g., Godfrey-Smith Reference Godfrey-Smith1994; Sokal and Rohlf Reference Sokal and Rohlf1994; Mayo and Spanos Reference Mayo and Spanos2006). Model fitting is generally fleshed out in terms of accommodating, rather than predicting, the evidence.Footnote ³ In evolutionary theory, model fitting has been articulated in terms of the general notion of empirical adequacy (e.g., Beatty Reference Beatty1980; Lloyd Reference Lloyd1983, Reference Lloyd[1988] 1994].

Models are often overfitted. That is, variables, parameters, and functions are altered or added in order to accommodate all of the available data. Lipton (Reference Lipton2005) calls this “fudging.” Hitchcock and Sober show that accommodation is easy. Moreover, they argue that “accommodational plasticity entails predictive impotence” (Reference Hitchcock and Sober2004, 7, 22): “When a background theory is sufficiently plastic that it can accommodate any data that may come along, it is in no position to make predictions about what data will come along” (Reference Hitchcock and Sober2004, 7). Overfitting occurs when accommodational plasticity is too high. This is a serious epistemic problem.

But some accommodation in the face of correcting evidence is necessary. Hitchcock and Sober (Reference Hitchcock and Sober2004) suggest an accommodation strategy that employs the Akaike Information Criterion (AIC). The AIC estimates predictive accuracy in terms of both the fit-to-data and the simplicity of a model (Hitchcock and Sober Reference Hitchcock and Sober2004, 12). Accommodation and prediction are intertwined in this account. However, in evolutionary theory per se the emphasis continues to be on (1) models that are confirmed and are accommodated to the evidence through regression analysis, curve fitting, and hypothesis testing, rather than on (2) the predictive capacity of models in new evidentiary contexts.

2.3. Predictive Capacity

There is a strong tradition in the philosophy of science arguing that the weight or relevance of a particular observation or piece of data for theory testing, confirmation, and choice is significantly greater if it is a novel prediction. A novel prediction meets the following conditions: (1) it was not known before the theory was constructed, or it was not used in the construction of the theory, and (2) the observation or data follows—deductively or even with high probability—from the theory (e.g., Popper Reference Popper1963; Lakatos Reference Lakatos, Lakatos and Musgrave1970; Zahar Reference Zahar1973; Gardner Reference Gardner1982; Giere Reference Giere and Earman1983; Howson Reference Howson1988; Worrall Reference Worrall, Gooding, Pinch and Schaffer1989, Reference Worrall, Cheyne and Worrall2006; and Laudan Reference Laudan1990). The view emphasizing the importance of novel predictions is sometimes referred to as “historicist” or “heuristic,” as opposed to the “logical” view (e.g., Musgrave Reference Musgrave1974; Mayo Reference Mayo1996; and Hitchcock and Sober Reference Hitchcock and Sober2004).

In contrast, the logical view holds that neither temporal order nor construction independence are epistemic issues directly relevant to theory justification (i.e., theory testing, confirmation, and choice). Formalist adherents of the logical view consider novel prediction to be completely irrelevant to theory justification (e.g., Hempel Reference Hempel1945, Reference Hempel1965; Carnap Reference Rudolf1950; see Musgrave Reference Musgrave1974). Other partisans are more forgiving. For screening-off advocates of the logical view, novel prediction is relevant to theory justification but only indirectly, as it gets philosophically screened off by more fundamental epistemic matters. Recent screening-off defenders have highlighted the following epistemic concerns: (1) severe testing through Neyman-Pearson, error-probing hypothesis testing (Mayo Reference Mayo1991, Reference Mayo1994, Reference Mayo1996, Reference Mayo2008; Mayo and Spanos Reference Mayo and Spanos2006) or (2) contrastive testing through either (a) maximum likelihood measures or (b) model selection criteria, such as the AIC (Forster and Sober Reference Forster and Sober1994; Sober Reference Sober1994, Reference Sober1999, Reference Sober2008; Hitchcock and Sober Reference Hitchcock and Sober2004). Although Mayo and Sober disagree on basic aspects of theory justification (e.g., Need it be contrastive? How important is it to evaluate and control for erroneous inferences of theory rejections and theory likelihoods?), they agree that novel predictions are not necessary.

The historicist view and the screening-off logical view agree that surprising, risky, and successful novel predictions, when they can be had, are impressive.Footnote ⁴ Under this cogent perspective, such predictions are sufficient for theory justification, though they are still not necessary according to screening-off philosophers. Predictive capacity (i.e., the ability to make surprising, risky, and correct novel predictions) is the theoretical virtue, and aim that I focus on in the case studies.

4.1. Parallel Evolution in the Bacterium E. Coli

Causal intervention in the sense of controlled and randomized experiments can be used to ground the predictive capacity of selectionist evolutionary theory. For 2 decades, Richard Lenski and coworkers have performed a remarkable set of experiments on the bacterium E. coli. In 1988, Lenski founded 12 populations of E. coli from a single ancestral population. Each day, the bacteria are transferred into new liquid media and are allowed to reproduce. They exhaust the glucose in the media after approximately seven generations (i.e., 8 hours). The bacteria then enter into a lag phase and “must wait until their 'springtime’ appears again the next day” (Lenski Reference Lenski2004, 226). The experiment has already surpassed 30,000 generations of bacteria.Footnote ⁵ The external conditions are controlled. The selection regimes are thus effectively isomorphic across the 12 lines. Evolutionary change can be measured in the laboratory.

Various kinds of parallel evolutionary changes have been observed. Each one of the 12 replicate populations has evolved larger-sized cells (Lenski Reference Lenski2004). Every population has lost the ability to catabolize most sugar sources other than glucose (e.g., each one is now unable to catabolize ribose; Cooper et al. Reference Cooper, Schneider, Blot and Lenski2001). These traits correlate strongly with fitness (Cooper and Lenski Reference Cooper and Lenski2000; Lenski Reference Lenski2004). Moreover, the genetic architecture underlying the traits is modified in impressively similar ways across the 12 populations (Cooper, Rozen, and Lenski Reference Cooper, Rozen and Lenski2003).

Parallel evolutionary change is readily predicted by selectionist evolutionary theory. In fact, parallel evolution in diverse lineages with a common ancestor is expected, and demanded, given similar selection pressures. This prediction is

1. 1. Risky: The absence of any significant parallel evolution would be a falsification of selectionist evolutionary theory.

2. 2. Surprising: Parallel evolution is not readily predicted under alternative theories/models such as historical contingency theory, which emphasizes random genetic drift (RGD) as an evolutionary force.Footnote ⁶

The novel prediction of parallel evolution in these experimental populations is a contrastive test of selectionist versus historical contingency theory. As Cooper and colleagues observe, “Parallel evolution of a trait across multiple lineages is often used as an indicator that the change is adaptive and has been shaped by natural selection” (Reference Cooper, Schneider, Blot and Lenski2001, 2834). Note also that the data were not used to construct the predictions of either theory. In short, defenders of the mechanism of selection should be relieved that selectionist evolutionary theory shows this sort of predictive capacity.Footnote ⁷

4.2. On the Endosymbiotic Origin of Eukaryotes

I now turn to a second type of case in which comparative (i.e., phylogenetic) inference is used in the context of selectionist evolutionary theory to formulate novel predictions. Causal-experimental manipulations are not possible here.

A scientific controversy during the '70s and '80s concerned the evolution of eukaryotic cells and their genome-containing organelles (e.g., mitochondria and chloroplasts) (Sapp Reference Sapp1994). There were two explicit alternatives:

1. 1. The endosymbiotic theory, which postulated the serial union, and uptake, of radically different sorts of prokaryotes over evolutionary time (e.g., Margulis Reference Margulis1970, Reference Margulis1975, Reference Margulis1976).

2. 2. A family of autogenous theories, which explained the internal origin of genome-containing organelles through the “pinching off” of membranes and subsequent intracellular sequestering of different internal genomes and metabolic networks (e.g., Raff and Mahler Reference Raff and Mahler1972, Reference Raff and Mahler1975; Uzzell and Spolsky Reference Uzzell and Spolsky1974; Cavalier-Smith Reference Cavalier-Smith1975; Reijnders Reference Reijnders1975; Taylor Reference Taylor1976).

Although both appealed to selection, they provided different explanations and novel predictions.

Let me be very specific regarding the predictions made. Margulis (Reference Margulis1975) provides an impressive list of 15 predictions inferred from the endosymbiotic theory. Natural selection is an explicit part of her framework (e.g., Reference Margulis1975, 23). She claims that while “several” of the “experimentally verifiable predictions are not absolute requirements of the theory, the phenomena listed here would be much more likely consequences of the serial endosymbiotic theory than other suggested models of eukaryote organelle origin (e.g., Raff and Mahler 1975)” (Margulis Reference Margulis1975, 28). Now, by observing that they are “much more likely consequences,” she certainly considers the novel predictions surprising. But by noting that they are not “absolute requirements,” she downplays their risky aspect. Even then, it is clear that four predictions were indeed risky: had they not been satisfied, Margulis's theory would have been significantly weakened, even falsified. Here are the novel predictions:

1. 1. Eukaryote transformation (Margulis Reference Margulis1975, 29). “The phenomenon of gene transfer from prokaryotes or eukaryotic donors to eukaryote nuclei will be demonstrated.” Compare: “the integration of the endosymbiont-proto-mitochondrion required wholesale transfer of genes from the endosymbiont genome to an unrelated nuclear genome. A mechanism by which this end may have been achieved is extremely difficult [!] to conceive” (Raff and Mahler Reference Raff and Mahler1972, 575–576). Margulis's surprising and risky novel prediction of gene transfer has been borne out; Raff and Mahler's hypothesis has been falsified (e.g., Adams and Palmer Reference Adams and Palmer2003; O’Malley and Dupré Reference O’Malley and Dupré2007).

2. 2. The red algal thallus was a multicellular heterotroph that acquired cyanelles (Margulis Reference Margulis1975, 28). The proto-red algae were heterotrophs. That is, they ingested other prokaryotes and did not synthesize their own energy sources (e.g., sugars). They then entered into selectively beneficial mutualisms with blue-green algae. Hence, the “metabolic pathways and primary amino acid sequences” of their “photosynthetic plastids” should resemble those of “blue-green algae” and not those of red algae themselves, as the autogenous theory predicted. This novel prediction has been amply corroborated (e.g., Oliveira and Bhattacharya Reference Oliveira and Bhattacharya2000). It confirmed endosymbiosis and helped falsify the family of autogenous theories.

3. 3. Hybridization between organelles and free-living micro-organism DNAs (Margulis Reference Margulis1975, 30). “Direct nucleic acid hybridization studies will show homologies between organisms and organelles as sketched in Fig. 1 [the reticulated phylogeny on p. 27].” This was the clincher novel prediction. Consider three different genomes: (1) eukaryotic nuclear DNA, (2) mitochondrial DNA, and (3) the DNA from a contemporary free-living prokaryote hypothesized to be related to the proto-mitochondria (i.e., a photosynthetic alphaproteobacteria). Endosymbiosis predicts that (2) and (3) will be sister taxa, with (1) as the outgroup. In contrast, autogeny predicts that (1) and (2) will be the sister taxa, with (3) as the outgroup. Impressive gene sequence analysis, which controlled for differential rates of evolution and other confounding factors, has amply corroborated the novel predictions of endosymbiotic theory (e.g., Yang et al. Reference Yang, Oyaizu, Oyaizu, Olsen and Woese1985; Gray, Burger, and Lang Reference Gray, Burger and Lang1999).

4. 4. Anastomosing phylogenies (Margulis Reference Margulis1975, 30–31). “A consistent phylogeny at higher taxonomic levels (such as is available for tracheophytes and chordates) acceptable to botanists, zoologists and microbiologists … will only be possible after acceptance of the symbiotic theory. … Because of anastomosing [reticulating] relationships … the finalization of such a phylogeny may be difficult.” Her surprising and risky novel prediction has been strongly confirmed. Endosymbiosis, and lateral gene transfer more generally, makes a neatly divisional, hierarchical tree of life difficult to draw.Footnote ⁸

In summary, Margulis's Reference Margulis1975 paper presents a number of key surprising, risky, and correct novel predictions. She relies on comparative data and comparative inference to express the predictive capacity of her theory. At least one of the 15 predictions remains undecided: eukaryotic flagella stem from a spirochaete symbiosis (Margulis Reference Margulis1975, 30). However, most have been strongly supported. The predictive capacity of the endosymbiotic theory, which relies on natural selection and on a novel way to think about descent with modification—that is, reticulation—is strong.