Grounding internalism in the population genetics of the introduction process

(work in progress: emerging manuscript, currently in the form of a persuasive essay)

To understand how to specify a causal grounding for internalist concerns, we must remove a series of obstacles, beginning with a major historic error in the use of population-genetic reasoning — an error whose effects reverberate today — concerning a possible causal link between internal tendencies of variation and tendencies of evolution.

The Haldane-Fisher argument and the SGFT

Futuyma (1988) attributes 3 remarkable accomplishments to an “Evolutionary Synthesis” of the 20th century: re-establishing neo-Darwinism on a Mendelian basis, sweeping away all rival theories, and providing a common framework for scientists in various disciplines to address evolution.

How were rival theories rejected? In an argument repeatedly cited by leading thinkers, Haldane (1927) and later Fisher (1930) concluded that mutation is a weak force unable to overcome the opposing pressure of selection, important only when selection is absent or when mutation rates are abnormally high (for more detail, see here). The argument was understood to mean that, given the observed smallness of mutation rates (and the equally well recognized pervasiveness of selection on visible traits), internalist theories relating tendencies of evolution to tendencies of variation are incompatible with population genetics, e.g., Gould (2002) writes as follows, citing Fisher (1930):

“Since orthogenesis can only operate when mutation pressure becomes high enough to act as an agent of evolutionary change, empirical data on low mutation rates sound the death-knell of internalism.” (p. 510)

This is a formal argument with a clearly recognizable logic, the effect of which is to reject an entire class of internalist theories, with no need for difficult experiments or time-consuming analyses of data! Accordingly, Provine (1978), in “The role of mathematical population geneticists in the evolutionary synthesis of the 1930s and 1940s,” identifies this argument as a key theoretical claim (see also Stoltzfus, 2017, 2019).

“For mutations to dominate the trend of evolution it is thus necessary to postulate mutation rates immensely greater than those which are known to occur.” “The whole group of theories which ascribe to hypothetical physiological mechanisms, controlling the occurrence of mutations, a power of directing the course of evolution, must be set aside, once the blending theory of inheritance is abandoned. The sole surviving theory is that of Natural Selection” (Fisher, 1930)

“For no rate of hereditary change hitherto observed in nature would have any evolutionary effect in the teeth of even the slightest degree of adverse selection. Either mutation-rates many times higher than any as yet detected must be sometimes operative, or else the observed results can be far better accounted for by selection.” (p. 56 of Huxley, 1942)

“If ever it could have been thought that mutation is important in the control of evolution, it is impossible to think so now, for not only do we observe it to be so rare that it cannot compete with the forces of selection but we know this must inevitably be so.” (p. 361 of Ford, 1971)

Figure: Some leading thinkers who invoked the Haldane-Fisher opposing pressures argument (clockwise from left: Haldane, Fisher, Huxley, Mayr, Simpson, Ford, Wright).

But the argument is wrong.

An evolutionary process that depends on events of mutation that introduce new alleles is subject to biases in mutational introduction, by a first come, first served dynamic that Haldane and Fisher did not address.

The flaw in the Haldane-Fisher argument arises from the assumption, common in classical thinking, that evolution can be treated as a process of shifting the frequencies of alleles in an initial “gene pool,” without events of mutation that introduce new alleles (see note 9). Of course, new mutations have to be involved somewhere in evolution, but they don’t have to be directly involved: if all the relevant mutations happened in the past, and the corresponding variant alleles are present in the gene pool at frequencies resistant to random loss, then we don’t need to address new mutations to understand evolutionary dynamics, which would follow merely from shifting gene frequencies.

Thus, the shifting-gene-frequencies theory (SGFT) posits that evolution can be understood as a shift from an initial multi-locus distribution of allele frequencies, to a final distribution of frequencies for the same alleles. This is what “evolution is shifting gene frequencies” meant, in practice.

That sounds surprising, right? Perhaps the skeptical reader will suspect that I am misrepresenting this issue or, at least, exaggerating. I get that. But it is important to establish this point, so I am going to make a blatant appeal to authority by quoting 3 completely independent sources —authored by eminent evolutionary geneticists — that note precisely this same restriction in classical population genetics:

“The process of adaptation occurs on two timescales. In the short term, natural selection merely sorts the variation already present in a population, whereas in the longer term genotypes quite different from any that were initially present evolve through the cumulation of new mutations. The first process is described by the mathematical theory of population genetics. However, this theory begins by defining a fixed set of genotypes and cannot provide a satisfactory analysis of the second process because it does not permit any genuinely new type to arise. ” (Yedid and Bell, 2002)

“Almost every theoretical model in population genetics can be classified into one of two major types.  In one type of model, mutations with stipulated selective effects are assumed to be present in the population as an initial condition . . . The second major type of models does allow mutations to occur at random intervals of time, but the mutations are assumed to be selectively neutral or nearly neutral.” (Hartl and Taubes, 1998)

“We call short-term evolution the process by which natural selection, combined with reproduction . . ., changes the relative frequencies among a fixed set of genotypes, resulting in a stable equilibrium, a cycle, or even chaotic behavior. Long-term evolution is the process of trial and error whereby the mutations that occur are tested, and if successful, invade the population, renewing the process of short-term evolution toward a new stable equilibrium, cycle, or state of chaos.” (p. 182). “Since the time of Fisher, an implicit working assumption in the quantitative study of evolutionary dynamics is that qualitative laws governing long-term evolution can be extrapolated from results obtained for the short-term process. We maintain that this extrapolation is not accurate.  The two processes are qualitatively different from each other.” (Eshel and Feldman, 2001, p. 163)

All three quotations suggest the same two things: (1) the SGFT was a limiting paradigm in mainstream 20th-century theoretical population-genetics, and (2) as the 20th century closed, this limiting paradigm was breaking down. The breakdown started, perhaps, with origin-fixation models in 1969. For many years, these models were used primarily with neutral or slightly deleterious mutations: this explains the oddly defined second category of Hartl and Taubes above. Eventually, SSWM models of adaptation (which overlap in meaning with origin-fixation models) emerged from Gillespie, and became the basis of the minor renaissance in modeling adaptation by Orr and others in the 1990s. That is, theoreticians have moved beyond the SGFT and embraced models of what is sometimes called the “lucky mutant” view or “mutation-driven” evolution, to such a degree that these models now represent a major branch of theory with diverse applications (see McCandlish and Stoltzfus, 2014).

But the SGFT was influential in the past, and remains cryptically influential today. Michod (1981) identifies a shifting-gene-frequencies paradigm as the “hard core” of a research program per Lakatos:

(a) The Hard Core

The basic elements of Lakatos’s model are all clearly identifiable within the population genetics research programme. For the population geneticist, the common denominator of all evolutionary forces is their effects on gene frequencies. In other words, gene frequency changes are evolution. This proposition, the hard core of population genetics, is best summarised by Sewall Wright in the conclusion to volume II of his treatise (Wright [1969], p. 472): 

“. . the species is thought of as located at a point in gene frequency space. Evolution consists of movement in this space.”

 This point of view is the basis of the population genetics approach to evolution. This is as true today as it was during the synthesis of the 1920s and 30s.

Michod is correct to think about this as a paradigm, because of the way it defines a broad and powerful perspective on how to think about the problems of evolution, answering basic questions that otherwise might be very hard to answer, and which might be answered quite differently by scientists working on evolution from different perspectives:

  • What is evolution? How do I know if it has happened?
  • Where does evolution take place? What is the causal locale?
  • How do I model evolution? What is the field or state-space?
  • What are the causes of evolution? How do I quantify them and weigh their importance in evolution?
  • How do I study evolutionary causes?

Certainly there were no agreed-upon answers to these questions prior to the Synthesis era. The distinctive answers suggested by the shifting-gene-frequencies paradigm shaped the Synthesis movement:

What is evolution? How do I know if it has happened? Evolution is shifting gene frequencies. Evolution has happened if there has been a shift in gene frequencies at the population level. A single event of birth or death is not evolution, and likewise, an event of mutation or recombination is not evolution. Instead, evolution has happened if there has been some significant shift in allele frequencies.

Where does evolution take place? Evolution takes place in populations because populations are the things that have allele frequencies. Individuals do not have allele frequencies. Species have allele frequencies, but only because they exist as populations (one or more) with allele frequencies.

How do I model evolution? What is the field or state-space? The population is the entity whose changes in allele frequencies constitute evolution. The course of evolution is the trajectory of a population in a field of allele frequencies. As Wright (above) suggests, “the species is thought of as located at a point in gene frequency space. Evolution consists of movement in this space.” A model of evolution represents the evolving thing as a population moving in its state-space of allele frequencies under the action of the forces.

What are the causes of evolution? How do I quantify them and weigh their importance? The causes of evolution are the processes that cause shifts in allele frequencies, in units of frequency change over time. The forces that cause larger shifts are, by definition, stronger forces.

How do I study evolutionary causes? The only direct way to study evolutionary causes is to focus on populations undergoing changes in allele frequencies, to assess what is causing those changes.

Some of the guidance provided by this paradigm turned into explicit dogma (e.g., the causes of evolution are forces that shift frequencies), and some of it was established more in the form of hidden assumptions or soft prejudices.

Considered more as a falsifiable claim than as a paradigm, the shifting-gene-frequencies theory asserts that we can understand evolution in nature adequately as a shift from one frequency distribution to another, so that any time-course of evolution can be represented as a trajectory in a continuous allele-frequency space.

(Figure legend) The shifting gene frequencies theory (SGFT). In the SGFT, adaptation — understood as a smooth shift in trait distributions (left) — is attributed to simultaneous shifts in the frequencies of alleles at multiple loci (middle), each with a small phenotypic effect. Formally, the population is a point in the topological interior of an allele-frequency space, i.e., the space of non-zero frequencies, and evolution is movement in this interior space (right). Thus, the forces of evolution are processes that can shift the population in this interior space. By contrast, the introduction process jumps a population from a surface into the interior where reproductive sorting processes (selection and drift) operate.

Within the SGFT, the forces of evolution are the biological processes that move the system in its state-space, i.e., the processes that shift frequencies. The ability to shift frequencies is obviously the measure of strength for a force: a biological process that causes larger shifts is necessarily a stronger force. Selection, being the strongest force, tends to dominate the process of shifting gene frequencies, i.e., it dominates the course of evolution.

What is the role of mutation in this theory?

In the process of shifting from an old to a new multi-locus frequency distribution, mutation pressure merely shifts the relative frequencies of pre-existing alleles. Because mutation rates are so small, these shifts are tiny in comparison to effects of selection and (typically) drift. Thus, the argument of Haldane and Fisher makes perfect sense within the SGFT.

That is, the Haldane-Fisher argument is both a fallacy and, at the same time, the correctly derived implication of the SGFT: if evolution can be adequately understood merely as shifting gene frequencies, then mutation is indeed a weak force, unimportant unless selection is absent (neutral characters) or the rate of mutation is unusually large.

Figure legend: The conclusion of Haldane (1927). The use of “results of importance” in the final sentence indicates that, for Haldane, the neutral case is not important, only the case in which the rate of mutation is large instead of small.

That is, mutation is a “weak force” in classical population-genetic thinking because the SGFT does not cover the novelty-introducing aspect of mutation. In effect, this aspect of mutation is treated as a background condition, rather than as a change-making causal process with explicit dynamics (see note 1). When a “gene pool” with pre-existing variation is assumed, the effect is that the novelty-introducing role of mutation is absorbed into this assumption as a background condition: the introduction process is literally is not part of “evolution” (shifting gene frequencies), but happens implicitly, before “evolution” gets started.

This theory makes mutation pressure largely irrelevant to modeling evolutionary change. This is why Lewontin (1974) says “There is virtually no qualitative or gross quantitative conclusion about the genetic structure of populations in deterministic theory that is sensitive to small values of migration, or any that depends on mutation rates.” The treatment of theoretical population genetics by Edwards (1977), shown in the image below, has hundreds of equations, but no terms for mutation. None. Not a single one. The word “mutation” appears only once in the entire book, on page 3, where the author says “All genes will be assumed stable, and mutation will not be taken into account.”

Note that the SGFT does not imply or suggest that new mutations never happen. Haldane, Dobzhansky and others stated explicitly that evolution ultimately would grind to a halt without new mutations. Instead, the verbal theory of the SGFT says that, even though new mutations are ultimately necessary, they are not immediately necessary, i.e., they are not directly involved, because the “gene pool” acts as a dynamic buffer, maintaining variation so that there is always abundant material for selection to respond to a change in conditions.

The popularity of the SGFT was driven partly by the sense that adaptation would be too slow if it involved waiting for the right mutation, instead of beginning with an abundant gene pool.

In addition, the SGFT was experimentally validated, a known mechanism. The experimental touchstone for the SGFT was Castle’s famous experiment with hooded rats (see Provine, 1971). Johannsen had already proven that selection is effective in sorting out true-breeding Mendelian types, but Castle and his colleagues showed something quite different. They started with a population of mottled black-and-white rats, and bred nearly all white, and nearly all black populations by selection in just 20 generations, not enough time for new mutations to play any appreciable role. This proved that selection could create “new types” (Provine) or “wholly new grades” (Castle) without the involvement of mutation, simply by shifting gene frequencies.

Finally, the SGFT provided a rhetorical foundation for Darwin’s followers to reject mutationism in the sense of “mutation proposes, selection disposes” (decides), a non-Darwinian theory distinct from their gradualist conception of evolution by the shifting and blending of abundant infinitesimals. The mutationist conception of evolution as a 2-step mutation-fixation process — the “lucky mutant” theory formalized in 1969 in origin-fixation models — is common today (read more here). However, the architects of the Modern Synthesis called on the SGFT to argue against the lucky mutant view (for more detail, see here). That is, even though the SGFT was a speculative theory, the architects of the Modern Synthesis convinced themselves that the theory was firmly established, and they conveyed this attitude of certainty to their readers, e.g.,

 “Novelty does not arise because of unique mutations or other genetic changes that appear spontaneously and randomly in populations, regardless of their environment. Selection pressure for it is generated by the appearance of novel challenges presented by the environment and by the ability of certain populations to meet such challenges.” (Stebbins, 1982, p. 160)

“It is most important to clear up first some misconceptions still held by a few, not familiar with modern genetics:  (1) Evolution is not primarily a genetic event.  Mutation merely supplies the gene  pool with genetic variation; it is selection that induces evolutionary change.” (p. 613 of Mayr, 1963)

This commitment continued to echo for decades in the notion that evolution does not depend on new mutations, a doctrine repeated in textbooks, e.g.,

“In practically all populations, however, the role of new mutations is not of immediate significance” (p. 464)

Strickberger MW. 1990. Evolution. Boston: Jones and Bartlett Publishers.

Thus, the SGFT was not merely a modeling convention — it was not just a technique used by mathematicians to make the equations easy to solve. Instead, the formal models and the conception of forces as mass-action pressures came together with a verbal theory about how evolution actually works in nature, and this integrated theory provided a basis to reject, not just orthogenesis, but mutationism in the sense of evolution via new mutations, i.e., mutation proposes, selection disposes (decides).

Even more broadly, the SGFT underlies the grand Synthesis claims noted earlier (Futuyma, 1988): restoring neo-Darwinism, sweeping away all rivals, and providing a unified framework for scientists in various disciplines to address evolution.

Yet, evolution in nature does not have to follow the SGFT. As stated earlier, an evolutionary process that depends on events of introduction — events of mutation that introduce a new allele, or events of mutation-and-altered-development that introduce a new phenotype — is subject to biases in the introduction process, by a simple “first come, first served” logic.

The logic of this theory was demonstrated by Yampolsky and Stoltzfus (2001) using a population-genetic model with 2 loci and 2 alleles. From the starting ab population, mutations with rates u1 and u2 introduce the beneficial genotypes Ab or aB, with a mutation bias favoring aB with magnitude B = u2 / u1 and with a greater fitness advantage (here, 2-fold) favoring Ab. The lines in the plot below all go up from left to right, indicating that the bias in outcomes (frequency of evolving aB relative to Ab) increases with the bias in mutation. The smaller populations show the degree of bias expected under origin-fixation dynamics (dashed line).

A distinctive prediction of this theory is that the influence of mutation biases does not require neutrality or high mutation rates (contra Haldane 1927), but will emerge (under the right conditions) from biases in ordinary types of nucleotide mutations, e.g., transition-transversion bias. This effect has been demonstrated conclusively in the past few years in both laboratory adaptation and in cases of natural adaptation (in diverse taxa) traced to the molecular level (for review, see Gomez, et al. 2020 or Stoltzfus, 2019).

Figure legend: The observed transition-transversion ratio among parallel adaptive changes is significantly higher than the null 1:2 ratio (Stoltzfus and McCandlish, 2017). This pattern is consistent with the theory of biases in the introduction process, but not with the mutation pressure theory of Haldane and Fisher.

Thus, a causal link between tendencies of variation and tendencies of evolution is theoretically possible and is actually observed. This result refutes a key argument from the mid-20th-century orthodoxy: internalist theories that attempt to link evolutionary tendencies to internal tendencies of variation are not inherently incompatible with Mendelian population genetics, but only with the SGFT (see note 2).


So far, we have established that the Haldane-Fisher argument is unsound theoretically, and that its conclusion is contradicted empirically. Haldane’s (1927) conclusion, even when considered narrowly, does not provide correct guidance for reasoning about evolution, e.g., when we see mutational patterns in molecular evolution, we cannot assume that this must reflect high mutation rates or neutral evolution. And the broad application of the Haldane-Fisher argument as a cudgel against internalism is wildly wrong.

Yet, in regard to the structure of evolutionary thought, much intellectual work will be required to reverse the damage done by this influential fallacy. Evolutionary discourse has proceeded through a century of theory development and exploratory thinking subject to the constraints that (1) a workable theory of biases in the introduction process was unknown to its major participants, and (2) the Haldane-Fisher argument placed a large “Do Not Enter” sign on the door leading to internalist thinking. This is a disturbing thought.

Do Not Enter Traffic Signs | Seton

This limitation was not known, for instance, in the 1980s, when the Modern Synthesis was being challenged on various fronts (molecular evolution, punctuation, evo-devo), and reformers were exploring new ways of thinking. Gould and Lewontin did not know it in 1979, when they wrote their famous critique of adaptationist thinking. Maynard Smith, et al. did not know it in 1985 when they wrote about “developmental constraints.” Kauffman did not know it in 1993 when, in The Origins of Order, he invoked “self-organization” to explain the findability of structures that are common in genetic state-spaces.

Yet all 3 sources are widely cited and have been influential — evidence of widespread hunger for internalist or structuralist alternatives to neo-Darwinism.

What happened, and what didn’t happen, because of this “do not enter” sign?

In the “spandrels” paper, Gould and Lewontin (1979) eviscerated the adaptationist research program, but their arguments for alternatives to natural selection were unconvincing.

Maynard Smith, et al. (1985), in their seminal piece on “developmental constraints,” noted explicitly that the Haldane-Fisher argument posed a barrier to the proposed efficacy of developmental biases in variation. If a theory of biases in the introduction process had existed in 1985, Maynard Smith, et al. could have used it to refute the Haldane-Fisher argument, and to provide a grounding for their claims regarding developmental effects, yet their foundational statement offers no general answer to the crucial problem of lacking a valid population-genetic basis (in the highlighted passage, they go on to suggest neutral evolution, obviously not an adequate foundation to address evo-devo concerns).

Accordingly, Reeve and Sherman (1993), in their subsequent defense of the adaptationist program, cited Gould and Lewontin as well as Maynard Smith, et al. (1985) and complained that the advocates of developmental constraint had offered no evolutionary mechanism. They call on the logic of the Haldane-Fisher argument when they ask, rhetorically, “why couldn’t selection suppress an ‘easily generated physicochemical process’ if the latter were disfavored?” Decades later, the notion of developmental constraint remains a flexible explanatory concept not tied to a specific evolutionary mechanism (see Green and Jones, 2016).

Yet, developmentalist-structuralist thinkers have continued to assume that internal factors actually matter in evolution, and because classical population genetics does not seem to provide a causal basis for this intuition, they have concluded that population genetics has some kind of metaphysical limitation that makes it inadequate as the basis for complete causal theories.

That is, population genetics is widely accepted as the language of causation in evolution, due to the influence of the shifting-gene-frequencies paradigm. Dobzhansky (1937) declared that “Since evolution is a change in the genetic composition of populations, the mechanisms of evolution constitute problems of population genetics.”   Yet, by the Haldane-Fisher argument, population-genetics rules out a dispositional role for internal variational factors. This has led internalist thinkers to suspect that something about population genetics makes it inadequate to construct complete accounts of evolutionary change.

The formalization of this complaint against population genetics is known as the causal completeness argument (Amundson, 2001). Because phenotypes exist and they are the stuff of evolution, an account of evolutionary causation that refers only to population genetics cannot be complete: development must fit in, somewhere, in a causal role. One way to integrate this role is to suggest that a full account of evolution must combine (1) the usual dry population-genetic account of causation by forces with (2) an alternative narrative of wet biological changes in development (e.g., Wilkins, 1998). This completes the causal account of evolution by supplementing standard forces with a kind of “lineage explanation” per Calcott (2009). In lineage explanation, the focus is on constructing a developmental-genetically plausible narrative for changes in a lineage over evolutionary time, as opposed to a focus on individual development over a lifetime, or on population genetics over evolutionary time.

In the more recent literature of EES advocacy, the diagnosis is converging on the conclusion that population genetics is too reductionistic, so that we need to “bring back the organism” as the focus of causal theories.

The problem of a missing causal foundation for internalism manifests differently in the (completely separate) literature of molecular evolvability or self-organization following on Kauffman (1993). What Kauffman needed was a theory explaining why certain features or forms emerge commonly by evolutionary processes, even without being selected (this was a big part of what Kauffman meant by the term “self-organization”).

A possible solution emerges from the fact that the structures that are more common in genetic state-space, e.g., RNA folds that have more possible sequences, necessarily have more mutational arrows pointed at them, including from other parts of state-space.

We might be tempted to suggest that this fact alone explains the findability of common structures, but this only tells us that a mutational bias exists — how such a bias influences evolution is a separate issue that requires a population-genetic theory linking tendencies of mutation to tendencies of evolution.

To grasp this point more clearly, think of Sober’s (1984) distinction of “source laws” and “consequence laws” of selection. Population genetics tells us how to compute what will happen in a population if A and B differ in fitness by some amount such as 2 %, given some background conditions including a scheme of heredity. That is, population genetics covers the consequence laws of selection. But it doesn’t tell where the differences in fitness come from, i.e., how they emerge biologically. For that, we need the source laws of selection, which come from physiology and ecology and so on.

Likewise, a complete causal theory for a variational influence would require both source laws that address how the variational tendencies emerge, and consequence laws that address their impact on evolution. As noted above, Maynard Smith, et al. (1985) drew attention to the source laws for developmental tendencies of variation, but failed to supply a consequence law linking those to measurable evolutionary effects.

More generally, in the evo-devo literature, the focus is on developmental source laws, and the issue of consequence laws is often not identifiable (e.g., here), so that the assumption that tendencies of variation must somehow cause evolutionary tendencies is wholly implicit.

By contrast, for evolutionary geneticists, the issue raised by evo-devo is precisely this alleged causal link between developmental biases and evolutionary ones, a link that is considered problematic and unlikely. For instance, in “Mutation predicts 40 million years of fly wing evolution”, Houle, et al. (2017) have done a superlative job showing a detailed quantitative correlation between (1) measured tendencies of varigenesis, i.e., new phenotypic variation M, and (2) measured patterns of evolutionary divergence R. But the authors themselves reject the theory that the pattern in divergence is caused by propensities of varigenesis. Quantitative genetics provides consequences laws relevant to this, but they are of limited scope and utility (see note 5), and the authors reject them for specific reasons.

When we are considering discrete traits, the Haldane-Fisher argument provides the consequence laws for biases in variation under the SGFT. Mutation is a weak force because mutation rates are small. Therefore, tendencies of mutation cannot be difference-makers in evolution, except in the case of neutral characters or unusually high mutation rates (Haldane, 1927).

Today, however, we can reject the SGFT and the Haldane-Fisher argument, and instead invoke the mutationist dynamics of origin-fixation models (for instance) to propose that the joint probability of origin-and-fixation of common structures (i.e., common in abstract genotype-spaces) is higher because their probability of mutational origin is higher. In this way, we can specify a complete chain of causation linking (1) a source law specifying that common structures have more mutational arrows pointed at them, with (2) a consequence law specifying that biases in the mutational introduction of alternative structures impose a bias on evolution (dependent on conditions).

But this kind of reasoning did not exist in 1993, and few scientists know about it today. Thus, proponents of effects of findability state the issue in other ways, e.g., Kauffman invoked “self-organization.” The general response of evolutionary geneticists to Kauffman’s work was that he clearly had some fascinating results, but it was not clear how relevant they were (given the abstractness of the models), or what they said about evolutionary causes. Kauffman repeatedly said that selection and “self-organization” worked together, in a partnership. But Kauffman was not calling on the usual list of evolutionary forces that shift allele frequencies to give a mechanistic account of self-organization, so we had no way to evaluate the causal status of this partnership. One reviewer called his references to self-organization “almost magical” (Fox, 1993).

In other parts of the contemporary literature on molecular evolvability, the effects of proximity and cardinality (of connected phenotype networks in genotype-space) that, in the above interpretation, are mediated by biases in the introduction process, are described as an effect of background conditions, as “constraints” emerging from properties of fitness landscapes (e.g., here), rather than being described in terms of causal forces.

Figure legend: Frequency vs. rank for the most common types of RNA folds of 100-nt sequences (from Dingle, et al, 2021). The circled folds are the ones found in nature. Thus, natural evolutionary processes discover the folds that are most common in sequence space.

In this way, the findability phenomenon is presented as something related to the complexity of the space in which evolution happens, i.e., patterns emerge, not due to any particular evolutionary force, but due to the unavoidable geography of the state-space for evolution. Yet the evidence that findability emerges specifically from the way that this state-space is sampled by mutation is suggested by studies such as Dingle, et al (2020): the findability effect (which they call “phenotype bias”) disappears when sampling compensates for the differing cardinality of structures in sequence space.

A causal grounding for internalism

“What the world most needs, then, is not a good five-cent cigar, but a workable — and correct — theory of orthogenesis.” (Shull, 1935)

Thus, the central barrier to establishing a causal grounding for internalist thinking in evolutionary biology is not that the domain of population genetics is too reductionistic.

Instead, the general problem is that the theory of causal forces is grounded specifically in a limited conception of population genetics (the SGFT), rather than in a more general conception of evolutionary genetics that depends on events of introduction.

In the conception of causation grounded in the SGFT, a causal force is a mass-action pressure modeled after the pressures of statistical physics, i.e., a pressure is a pressure on allele frequencies, and it results from aggregating over the effects of innumerable events among the individual member organisms of a population (see Sober, 1984). Selection and drift result from the aggregate effects of innumerable births and deaths. The force of mutation is mutation pressure, the aggregate effect of innumerable events of mutational conversion in different individuals.

Just as statistical physics is not a reductionist theory, the SGFT is simply not a reductionist theory. The SGFT clearly posits an emergent population “level”, and the architects of the Modern Synthesis argued explicitly that the forces of evolution are emergent at the population level, and do not exist at the more reduced level of individual organisms. For instance, in their textbook, Dobzhansky, et al (1977) write

Each unitary random variation is therefore of little consequence, and may be compared to random movements of molecules within a gas or liquid.  Directional movements of air or water can be produced only by forces that act at a much broader level than the movements of individual molecules, e.g., differences in air pressure, which produce wind, or differences in slope, which produce stream currents.  In an analogous fashion, the directional force of evolution, natural selection, acts on the basis of conditions existing at the broad level of the environment as it affects populations. (p. 6)

An individual event of mutation that introduces a new allele does not satisfy this conception of an evolutionary cause. It is a proximate cause, in the language of Mayr. Likewise, the development of an individual is a proximate cause.

Another way to state this limitation is that the prevailing theory of causal forces used in evolutionary thinking recognizes causes of fixation but not causes of origination, when in fact a full account of evolutionary causation requires that both origination and fixation are treated as change-making causal processes with explicit dynamics.

The flaw in the forces theory is exactly the same thing as the flaw in the SGFT. The sufficiency of the SGFT depends on evolution remaining in the topological interior of the relevant allele-frequency space, where all frequencies are non-zero. In this topological interior, all of the classical forces are identical in the sense that each force can change a frequency f to f + δ, where δ is an infinitesimal. This common currency of causation is how the scheme of “forces” achieves generality. In the interior, any infinitesimal change, anywhere in the space, can happen by any force. This means that we can chain together any series of infinitesimal changes into a trajectory, and this trajectory can (in principle) be caused by any force, or by any combination of forces.

But this scheme of causation by forces falls apart if we consider movement from the surface (edge) of an allele-frequency space into the interior. In the left figure below, we have a small shift from the center of a 2-dimensional allele-frequency space (0.500, 0.500) to (0.502, 0.506). This could be caused by selection, drift or mutation, or by any combination of them, although a shift of 1 % is enormous for mutation alone, and would not normally happen in one or a few generations: mutation is indeed a weak force. The right figure shows a shift from (0.000, 0.000) to (0.002, 0.006) that is mathematically identical, but not evolutionarily identical, because mutation is absolutely required to cause a shift upward from 0.

The logic of forces breaks down because the impact of the biological process of mutation is weaker than every other force in the interior of an allele-frequency space, but infinitely stronger than every other force at the surfaces, where it acts by discrete events, not continuous shifts. This is qualitatively different causal behavior. When an evolutionary process includes discrete events of introduction that jump an evolving system off of the surface of an allele-frequency space into the interior (where the forces of selection and drift operate), mass-action pressures are not a sufficient guide to causation.

This argument might sound very abstract, thus not relevant to practical evolutionary reasoning. Probably some readers, at this point, are cursing at their screens and screaming “Of course we recognize the introduction process as part of evolution!!! We’re not a bunch of idiots!!” And my response to that is yes, of course you recognize that evolution must depend on events of mutation that introduce new alleles, and of course you may have used origin-fixation models or methods of analysis that assume origin-fixation dynamics, and perhaps if you work in molecular evolution, you are a full-on mutationist who habitually thinks about evolution as a two-step proposal-acceptance process. But that is not the same thing as recognizing the introduction process formally as a kind of causation distinct from classical pressures, because if you have been thinking that way already, you would not be cursing at your screen, because you would already understand how this changes everything.

Anyone who saw how the evo-devo challenge played out in the 1980s and 1990s knows that abstract arguments about what qualifies as a true evolutionary cause (i.e, not development) have been deployed with great effect against claims of novelty from evo-devo. For instance, Wallace (1986) asks whether embryologists can contribute to understanding evolutionary mechanisms, and then answers negatively, arguing that “problems concerned with the orderly development of the individual are unrelated to those of the evolution of organisms through time.”

“If we are to understand evolution, we must remember that it is a process which occurs in populations, not in individuals.  Individual animals may dig, swim, climb or gallop, and they also develop, but they do not evolve.  To attempt an explanation of evolution in terms of the development of individuals is to commit precisely that error of misplaced reductionism of which geneticists are sometimes accused” (Maynard Smith, 1983, p. 45).

“I must have read in the last two years, four or five papers and one book on development and evolution.  Now development, the decoding of the genetic program, is clearly a matter of proximate causations.  Evolution, equally clearly, is a matter of evolutionary causations.  And yet, in all these papers and that book, the two kinds of causations were hopelessly mixed up.” (Mayr, 1994) 

“No principle of population genetics has been overturned by an observation in molecular, cellular, or developmental biology, nor has any novel mechanism of evolution been revealed by such fields.” (Lynch, 2007)

To escape this kind of smack-down, presumptive causal arguments from evo-devo, or from any other sub-field in evolutionary biology, must refer to the forces of population genetics, because the statistical forces theory is the only accepted theory for what constitutes a genuine evolutionary cause.

But if events of introduction are evolutionary causes, and biases in the introduction process are causes of evolutionary bias, then

  • events of mutation can be evolutionary causes,
  • events of mutation-and-altered-development that introduce new phenotypes can be evolutionary causes, and
  • mutational and developmental biases in the generation of variation can be evolutionary causes.

For instance, recognizing the introduction process provides the place to recast the plausibility arguments in lineage explanation into arguments about the developmental factors that induce biases in the introduction process.

That is, the solution to finding a causal grounding for internalist thinking is not to repudiate population genetics or supplement it with a parallel plane of developmental causation, but (1) to point to the part of mathematical theory covering the introduction process, which already exists and is currently in active development, and (2) taking account of what we now know about the powerful influence of this qualitatively distinct process on the observed course of evolution, and the theoretically expected course, finally (3) to insist that we must locate a cause in this part of population genetics, i.e., we must declare that the introduction (origination) process is a genuine cause, a change-making causal process with propensities that must be treated explicitly (again, see note 2).

To be perfectly clear, the necessity of doing this, justifying the use of “must” in the previous paragraph, is that the evidence cited above compels us to recognize that the dynamics of mutational introduction are profoundly consequential, and our theorizing suggests that a far broader role is inevitable. The notion that we can treat evolution as shifting gene frequencies, without directly involving the dynamics of introduction, has been rejected, and the required correction to our conception of causation is to recognize the introduction process (mutational or otherwise) as something that must be treated explicitly as a cause in order to get evolution right.

[The last two paragraphs are worth re-reading. ]

From the beginning, critics of Darwin’s thinking have objected that selection does not create anything new. Darwin’s followers developed several well known responses to this objection, justifying the creativity of selection. One of them is essentially that there is infinitesimal variation in every trait, and selection can leverage that diversity to create novelty by quantitative shifts. In world of continuous quantities, this is abstractly true. Another response is that selection is creative in the sense of bringing together rare combinations out of the diversity of the gene pool. This is also true, in a sense (dependent on recombination). Another argument is that selection can accrue effects in a particular direction, consistently, over long periods of time. This, too, is clearly true.

And yet, ultimately, these justifications of the creativity of selection do not address the fundamental problem that we cannot get the dynamics of evolution right without representing discrete events of the introduction of novelty by mutation-and-altered-development. We must recognize the introduction process as a genuine evolutionary cause.

Once we have added this vital piece of conceptual infrastructure, it then becomes possible to build a larger framework for causal theories. By appealing directly to the introduction process as an alternative type of population-genetic causation, we can specify complete chains of causation from internal features that determine mutational and developmental propensities of variation, to quantifiable evolutionary behavior, via the population-genetic consequences of biases in the introduction process.

Let us briefly consider how to utilize the concept of biases in the introduction process (as a genuine cause of evolutionary orientation or direction) to specify a causal grounding for 3 historic themes of internalist thinking:

  • Taxon-specific propensities
  • Intrinsically likely forms
  • Directional trends

The first step is to make the transition from mutation biases and nucleotide-level effects to phenotypes. Let varigenesis cover all of the processes involved in the generation of new variation, from mutation to altered phenotypic development, subject to any applicable conditions. In quantitative genetics, varigenesis is represented by the M matrix of variances and covariances for new phenotypic variation (see note 5). For a discrete phenotype-space, we may consider a vector of rates U, with one rate for each alternative phenotype.

In the original Yampolsky-Stoltzfus model, the mutation bias B is a ratio of two mutation rates u1 and u2 , and these are specific mutation rates from one genotype to another.

But when we turn our focus to phenotypes, we can simply redefine u1  and u2 in terms of alternative phenotypes. For instance, consider an example in the figure below, based on the genetic code, which is the genotype-phenotype (GP) map relating codon genotypes to amino acid phenotypes. The rate u1 represents Asp-to-Val and the rate u2 represents Asp-to-Glu, which implicates 2 different mutational paths. Therefore, even if all mutation rates are the same (no mutation bias), the GP map induces a 2-fold phenotypic bias favoring Asp-to-Glu.

To the extent that fitnesses depend only on the phenotype, the two mutational paths from Asp to Glu are identical and will behave as if this were 1 path with a 2-fold higher rate. In this way, all the same conclusions that apply to a B-fold bias in the Yampolsky-Stoltzfus model will also apply to a B-fold phenotypic bias in varigenesis. That is, a GP map such as the genetic code induces asymmetries in the introduction of alternative phenotypes.

The figure above right represents a precisely analogous idea that is common in the evo-devo literature, which is that some alternative phenotypes may be more likely (in varigenesis) because they implicate a larger number of mutationally accessible genotypes, i.e., genotypic neighbors in the GP map. Here, the 1-mutant neighborhood of a genotype encoding phenotype P0 includes 5 genotypes with phenotype P2, and only 1 with genotype P1. In this case, without any mutation bias per se, there is still a 5-fold phenotypic bias in varigenesis toward P2.

In general, the form of the above argument is to use the neighboring phenotypes implicated by a GP map to define equivalence classes of genotypes, so that we can aggregate mutation rates by equivalence class, with the result that the differential mutational accessibility of alternative phenotypes will emerge due to asymmetries in the GP map, even if all mutation rates are the same. That is, a mutation spectrum at the genotypic level, together with a GP map, induces a description of potentialities or dispositions of phenotypic change in a developmental-genetic system, i.e., a description of varigenesis.

This provides a rigorous justification for the notion that each taxon, having a distinctive GP map, has an intrinsic evolutionary potential or inherited predisposition, due to propensities of varigenesis.

Next, let’s consider the tradition of structuralist arguments to the effect that certain familiar structures or features commonly emerge in nature because they are, in some sense, intrinsically likely, i.e., because they are the most natural or easily generated states of the materials in question.

We already addressed the contemporary form of this argument per Kauffman, which has been made in regard to RNA folds, protein folds, regulatory network structures, and some features of tissue layers: the forms that are intrinsically likely are understood to be the forms that are common in genetic possibility-spaces, and the question of evolutionary causation is what evolutionary cause makes intrinsically likely forms evolutionarily likely.

In regard to RNA folds, the folds with the most sequences occupy the greatest volume in genotype space, thus they have the largest number of mutational arrows pointed at them, including the arrows pointed at them from other regions of genotype space (which is actually a function of surface area rather than volume or cardinality). This means they are more likely to be proposed, thus more likely to be proposed-and-accepted, by an evolutionary process that explores sequence space via mutation.

Figure legend: Two effects of mutational phenotype accessibility are inevitable when considering phenotype networks in genetic state-space (in the figure, mutation only samples adjacent vertices and each network represents genotypes with the same phenotype or fold). The shorter-term effect of mutational accessibility is that, from P0, P2 is more accessible than P1. The longer-term effect is that P0, with more genotypes, has a larger contour length (or more generally, surface area), thus it has more mutational arrows pointed at it from external regions of state-space.

Thus, it is possible to specify a rigorous causal grounding for the kind of structuralist argument that explains what is evolutionarily likely by referring to what is common in abstract possibility-spaces.

Next, consider the idea of long-term directional trends due to internal biases in variation. Classical thinking says that such trends are impossible, and that (except under the case of neutral evolution) selection is the sole source of direction in evolution. However, models of adaptive walks with protein-coding genes subject to GC or AT biases in mutation show that compositional trends are the predictable result of biases in the introduction process.

Figure legend: Mean trajectories of simulated adaptive walks on a protein NK landscape from Stoltzfus (2006), where evolution is subject to AT:GC mutation bias of 1:10 (blue), 1:3 (brown), 1:1 (green), 3:1 (orange) or 10:1 (red). Proteins adapting under AT bias evolve to be enriched for amino acids with AT rich codons (FYMINK), whereas those adapting under GC bias are enriched for amino acids with GC-rich codons (GARP).

Note that when we collapse evolutionary change down to 1 dimension, the result is that internal and external factors, if they do not coincide in direction, must clash in direction. This way of combining the two types of causes leads to a consideration of which force is stronger, and (given the “pressure” conception of forces) selection is assumed to be the winner of this zero-sum game. But for an evolutionary process operating in a high-dimensional space such as a protein fitness landscape, there are typically many directions toward increased fitness, some of them more favored by mutation, and some less favored. As a result, the trajectory of adaptive evolution in a high-dimensional space may have components of direction that are due to fitness effects, and other components of direction that are due to internal variational biases.

Thus, it is possible to specify a rigorous causal grounding for the notion of trends due to internal biases, and we can use this theoretical foundation to rebut the false intuition, widespread in the literature, that strong selection must necessarily suppress the effect of internal tendencies of variation, or that evolution is some kind of zero-sum game in which mutational tendencies have a necessary cost. From modeling, we know exactly why it is a mistake to think of mutation bias as a cost: the same mutation bias will make adaptation easier in some cases and harder in others, depending on circumstances, a point that is illustrated by Cano and Payne (2020) using empirical fitness landscapes.

To summarize, the paragraphs above outline a causal grounding for the classic internalist-structuralist themes of (1) taxon-specific propensities, (2) intrinsically likely forms and (3) directional trends. To suggest this causal grounding does not mean that all past internalist statements are true or even that they are all theoretically possible. Many of these past claims could be stupid. What it means is that, for each of these three classic types of claims, we can map the form of the claim onto a causal model that validates its logic. If we can map a specific internalist claim to a causal model of this form, then it becomes a substantive falsifiable hypothesis about internal causes that can be tested using whatever tools are available to test hypotheses.

“Adaptation has a known mechanism: natural selection acting on the genetics of populations … Thus we have a choice between a concrete factor with a known mechanism and the vagueness of inherent tendencies, vital urges, or cosmic goals, without known mechanism.” (Simpson, 1967, p. 159)

This means that we are in a different place than in 1967 when Simpson dismissed the notion of internal trends. Simpson’s argument is invalid, and this is not because we have discovered vital urges or cosmic goals, but because we have reconsidered evolutionary genetics both in theory and in fact, and we have concluded that internal biases are a real possibility grounded in the theoretically and empirically demonstrated effects of biases in the introduction process. The influence of such biases is now in the “known mechanism” category, available to be applied in all areas of evolutionary research.

Distinguishing other theories and paradigms

This line of argument for a rigorous basis for internalism does not map in a simple way to the current reformist literature in evolutionary biology, with its complaints about reductionism, calls for the “return of the organism,” and exploration of the diffuse EES-SET axis of dispute. The argument here evokes a very precise conflict with classic thinking, because classic thinking about the role of mutation depends for its validity on a specific conception of causal forces as mass-action pressures. The mapping to EES concerns would be clearer if the EES Front had defined itself more narrowly in opposition to the sufficiency of the shifting-gene-frequencies paradigm per Michod above.

Relative to the classic conception of evolutionary causes as population-level pressures on allele frequencies, the introduction process conflicts with the statistical pressure criterion, but not necessarily the population-level emergence criterion. The introduction process is arguably emergent at the population level: if a specific individual in state A1 mutates to state A2, this is clearly an event of mutation, but we cannot determine whether it is an event of introduction without examining the population of which the individual is a member, i.e., we can’t diagnose an introduction event except at the population level. Again, mutational introduction and mutational conversion are distinct: introduction is emergent at the population level (see note 7).

However, the introduction process is different from classical forces because it is not a deterministic mass-action pressure aggregating over the behavior of countless individual members of a population (see note 3). To characterize the introduction process as a cause is to put the focus on a probability distribution for events that reflect generative processes acting inside organisms.

The main distinction from the “constraints” literature of evo-devo is the concern to specify complete chains of causation from internal features that determine propensities of variation, to quantifiable evolutionary behavior, via population genetics. As explained above, the typical approach in the evo-devo literature, following Maynard Smith, et al (1985), leaves a gap in this chain of causation, where the missing theory would explain how developmental propensities of varigenesis become evolutionary propensities. In the evo-devo literature, efforts to reform thinking about causation typically focus on supplementing population-genetic causation with lineage explanation, rather than rethinking population-genetic causation.

The main distinction from the “evolvability” literature and the more recent literature of “developmental bias” in the EES context is that the causal grounding for internalist thinking offered here does not, in any way whatsoever, presume or imply that variation is facilitated, contrary to the fatuous treatment by Svensson and Berger (2019). The literature has been (to my way of thinking) relentlessly confusing on the extent to which the distinctiveness of evo-devo, or the distinctiveness of evolvability claims, is presumed to rest on facilitated variation, but the proposal here is perfectly clear: there is no necessary dependence.

Instead, the focus here is on consequence laws— consequence laws that apply whether or not any source laws exist that specify facilitated variation. For instance, it is not necessary to assume that the molecular bias for transition over transversion mutations is in some way beneficial (see Stoltzfus and Norris, 2015). The theory predicts an influence of transition bias on adaptation even when the mutation bias is perfectly orthogonal to fitness effects. The issue of whether varigenesis is dispositional in its effect on evolution can be adjudicated entirely separately from whether varigenesis is facilitated.

Of course, non-orthogonality is inevitable in a high-dimensional world. In the case of any real-world landscape, the mutational bias toward transitions (1) will tend to align the overall process of evolutionary exploration better (quantitatively) with beneficial trajectories, or (2) will tend to align it worse. The modeling study by Cano and Payne (2020) demonstrates this point using empirical fitness landscapes for binding sites.

Perhaps this will seem disappointing for those familiar with the literature of evolvability or the EES Front, where the idea that variation is facilitated, and that living systems are surprisingly “innovable,” is deeply entrenched. Where is the mojo of internalism if internal variational propensities are merely arbitrary and not an expression of the superior evolvability of naturally evolved systems?

My suggestion, in this context, is that we must learn to walk before trying to run. Perhaps variation is facilitated, but isn’t a dispositional evolutionary role for varigenesis both the premise and the promise of this claim? Isn’t it the first-order effect that necessarily underlies all possible higher-order effects pertaining to intrinsic variability?

In fact, the notion of taxon-specific propensities of variation that are merely dispositional without being facilitated is historically part of evo-devo, e.g., this was the explicit position of Maynard Smith, et al (1985). Similarly, the primary argument of Alberch and Gale (1985) is that two different taxa (salamanders and frogs) tend to lose digits differently in evolution (pre- or post-axially) because they tend to lose them differently when development is perturbed: Alberch and Gale were not arguing that each taxon evolved a better way to lose digits, but merely a different way.

What victory looks like

We may debate forever what a scientific theory is, but it is relatively clear what a scientific theory does: it serves as a guide to reasoning, a provisional domain-specific extension of logic. Theories tell us how to think about something, given a certain set of inputs and assumptions. Just as for a pasta machine, you give it some inputs and turn the crank, and you get some outputs with a shape that reflects the internal workings of the theory. The inputs are conditions, and the outputs are expectations or provisional hypotheses or explanations.

The theory could be wrong if the internal logic is mistaken, or it could be wrong if its outputs do not match reality.

Recently I was involved in discussion of an idea from the “directed mutations” literature called the “non-specific reverse-transcription” mechanism or NSRT, in which a beneficial error in an RNA transcript allows a starving cell to survive in a way that leads to reverse-transcription of the error back into the genome, resulting in a heritable mutation. This is a hand-waving theory, with an unvalidated logic. It specifies a mechanism for an unorthodox kind of mutation, but no one knows for sure whether a system with the parts and operations specified by the theory would actually have the beneficial-mutation-enhancing behavior specified by the theory.

This is why formal modeling is so important. In formal philosophical discussions, a model is a system of things that makes all the statements in a theory true. If someone could make a model — either an engineered biological system or a mathematical (or computational) model— that exhibits the proposed NSRT mutational behavior, this would validate the theory’s internal logic.

The theory of biases in the introduction process is not a hand-waving theory, but refers to a known mechanism of evolution. To say that this theory (in context) provides a causal grounding for internalism means that each of the classic internalist themes has a form for which a valid basis exists. The extent to which this possible causal basis is the actual causal basis for an influence of internal factors in evolution remains to be determined by appropriately designed research projects.

This research, of course, will require many people and many years.

In an ideal world, scientists would study the line of argument above and consider it seriously as a major unifying conjecture with broad implications for evo-devo, molecular evolution, population genetics, paleontology, and so on. Interested scientists would take the time to understand these ideas, and then begin the process of applying, evaluating, and so on.

Alas, the real world is not like that. Evolutionary discourse is like a meat grinder that mashes ideas into an inchoate mass (e.g., this), doing this with so much noise and splatter that sensible people step back and cover their ears. Even the scientists who are sympathetic to this line of argument will mangle it. And nothing brings out the haters more than claims of novelty.

To some extent, this is true in every discipline — new ideas generate a vitriolic and ignorant backlash (e.g,. think of the Neutral Theory). Once, long ago, I pulled out an old bound volume of Nature to read the original proposal of chemiosmotic coupling by Peter Mitchell. Mitchell (1961) proposed a role for membrane potentials that is now widely accepted, but which, at the time, challenged the conventional wisdom that energy storage and transfer is mediated by high-energy bonds in diffusible molecules. Mitchell’s hypothesis does not sound particularly dramatic or offensive. Who could possibly be upset about it? But in the copy of Nature 191 in the stacks at the Kellogg medical library at Dalhousie University, there was literally a “f*** off!” scrawled in the margin of Mitchell’s article (see note 8).

A proposal as broad as a causal grounding for internalism, calling on a new kind of causation and connecting threads from evo-devo, molecular evolution, self-organization, etc., is going to be mutilated. The haters will conflate it with other ideas that they hate, attacking it with arguments used against previous targets of their antipathy. The sympathizers will blend it and conflate it with their own favorite ideas, to make it more familiar, thus missing the point of what is new. There will be constant attempts to equate or associate this particular theory or line of argument with some other idea, either to appropriate it or to dismiss it (e.g., “this is is just X” where X is constraints, contingency, mutation pressure, new-words-for-old-concepts, population genetics, etc).

Figure legend: A gravity model for misapprehension and distortion. Gravitational pull on the focal object from a second object is a function of mass / distance2. In the chaos of evolutionary discourse, the theory proposed here will be distorted by, and pulled in the direction of, other ideas that, though clearly non-identical, have vague similarities and currently have a far larger mass (level of scientific activity, familiarity, commitment). Svensson and Berger (2019) illustrate a variety of ways to misrepresent the theory.

In this way, the sausage grinder of evolutionary discourse will mangle the above proposal unrecognizably (e.g., this), so that the vast majority of scientists who are exposed to representations of the theory will perceive it incorrectly. In particular, the idea will be pulled relentlessly in the direction of facilitated variation or directed mutation, both by haters who hate these ideas, and by sympathizers who favor them. The only hope for success is that a small group of open-minded scientists gets a clear grasp of the theory and builds up a body of work that eventually — if the theory is indeed powerful — becomes too substantial to ignore. Then, others will be forced to integrate the theory, and eventually it will be so well established that the fundamentalists from Answers in Synthesis will try to claim it for tradition.

The way to block out the noise and maintain a clear view of issues of science and history is to focus on theories (not words or people or flexible traditions or research programs) understood operationally by how the theories guide scientific reasoning, as documented by recurrent, explicit, and unmistakeable claims in the published literature (i.e., the approach used above). As noted, a genuine theory is a reasoning machine: provide an input and turn the crank, and the theory returns an output (a hypothesis, an expectation, a provisional explanation) that reflects the inputs and the logic of the theory. If the reasoning leading to the theory’s distinctive implications is absent in some source document, then the theory is absent, even if the source uses some of the right words (for an interesting edge-case, see note 6).

Which theories are (or were) actually used in reasoning about the role of variation in evolution? What kinds of reasoning do these theories support, as documented by recurrent and explicit claims in the evolutionary literature? That is, if we point to something about evolution, some outcome or pattern or phenomenon, and ask an evolutionist (or ask a historical source) what roles for variation have been considered explicitly in evolutionary accounts of that outcome or pattern, what consistent answers do we find? Here are some:

  1. Variations emerged adaptively by effort, and were preserved, as per Lamarck
  2. Variation supplied indefinite raw materials that selection shaped into adaptations, as per Darwin.
  3. The mechanisms of development imposed constraints, setting limits on what is possible, as per Eimer (1898) or Oster and Alberch (1982)
  4. Mutation pressure drove allele frequencies to high levels, under neutrality or high mutation rates, per Haldane (1927)
  5. New quantitative variation (M) contributed to standing variation (G) which, together with selection differentials (β), jointly determined (as ) the short-term rate and direction of multivariate change in quantitative characters (Lande and Arnold, 1983; see note 5)
  6. Mutational and developmental biases in the introduction of variation imposed biases on evolution by a first come, first served logic, without requiring neutrality or high mutation rates (Yampolsky and Stoltzfus, 2001)

The first 4 ideas appear repeatedly in the classic literature. What does each theory say, or not say, on the relationship between tendencies of variation and tendencies of evolution? How does each theory support (or not support) a causal basis for internalism?

The authors of Maynard Smith et al. (1985) — including Pere Alberch (theory 3), Russ Lande (theory 5), Stuart Kauffman (mentioned above in regard to self-organization) and John Maynard Smith, who studied with Haldane (theory 4)— each had access (in principle) to all of the theories that existed in 1985 in the literature. They cited the Haldane-Fisher mutation pressure theory (theory 4). They cited Vavilov’s (1922) informal theory of parallel evolution due to parallel variation. Their discussion clearly considers the potential influence of merely quantitative biases (as distinct from hard limits) in phenotypic variation due to the influence of developmental mechanisms on varigenesis (image below).

Figure legend: Here Maynard Smith, et al (1985) are clearly talking about biases in phenotypic variation due to the influence of developmental mechanisms on varigenesis, where the biases are quantitative (“more readily accessible”) rather than absolute, and they are not assumed to be facilitated or adaptive.

Yet, as noted above, available scientific knowledge did not equip these leading thinkers with the conceptual or theoretical tools to rebut the Haldane-Fisher argument and offer a formal basis to interpret developmental biases as dispositional evolutionary causes. The last theory listed above, the one about biases in the introduction process as a cause of orientation or direction in evolution, simply was not available in 1985.

In fact, a clearly worked out theory of biases in the introduction process is conspicuously absent from the 20th-century literature bearing on (1) the Haldane-Fisher argument that internalist theories are incompatible with pop-gen; (2) the assumption, common in the molecular evolution literature since Sueoka (1962), that mutational explanations for patterns somehow require or imply neutrality; (3) the case of developmental constraints discussed above; (4) the case of findability discussed above; (5) the notion of trends due to speciation bias or directed speciation that has been floating around in the macroevolution literature since the 1970s (see note 6).

Figure legend: a brief (15 min) talk pointing to 4 topics in evolutionary discourse where a theory for the efficacy of biases in the introduction process is (retrospectively) conspicuously absent.

The theory of biases in the introduction process as a cause of orientation or direction in evolution, again, is a specific theory with distinctive and testable implications. The logic of the theory generates various outputs that are otherwise not known to be part of evolutionary reasoning. Indeed, one way to explore this distinctiveness is to use the rhetorical approach of crafting statements that the theory distinctively enables. Such statements can refer, not only to expected evolutionary behavior, but also to other theories, and to informal claims in the literature that may be supported or contradicted, like these:

  • Biases in the introduction of variation can impose biases on the course of evolution without requiring neutrality, high mutation rates, or absolute constraints
    • Thus, variational biases on the course of adaptation are possible.
    • The common assumption in the molecular evolution literature that mutational effects on evolution require or imply neutrality is mistaken
    • The Haldane-Fisher argument as expressed by Haldane (1927) or Fisher (1930), and as employed by authors such as Huxley, Ford, Gould, Maynard Smith, et al, is not valid
    • The joint dependence (that emerges under some conditions) of adaptive changes on fixation probability and chance of mutational introduction invites previously unimagined considerations of Berkson’s paradox
  • For moderate values of B, there are conditions (e.g., in the origin-fixation regime) under which a B-fold bias in the introduction of variants results in a B-fold bias in evolutionary change
  • Biases in mutational accessibility of alternative phenotypes represent a kind of developmental bias, and conditions exist under which this kind of developmental bias may influence evolution in the same way, i.e., by the same kind of population-genetic mechanism, as a mutational bias of the same magnitude
    • This result invalidates the historically important argument (by Mayr, Wallace and others) attempting to undermine causal claims of evo-devo on the grounds that development cannot be construed as an evolutionary cause.
  • Adaptive traverses of high-dimensional spaces can exhibit, simultaneously, components of direction that reflect fitness effects mediated by selection, and components that reflect biases in varigenesis mediated by the introduction process
    • This result invalidates a kind of informal logic (of selection as a governing force) suggesting that internal biases must come at an adaptive cost or that they somehow work against or impede selection
  • Systematic biases in the mutational introduction of phenotypic forms (due to their differing surface area in genotype-space) provide a possible population-genetic mechanism for the findability aspect of “self-organization” reported by Kauffman (1993) or the “phenotype bias” reported by Dingle, et al (2021).


The notion that evolution may have tendencies that reflect internal tendencies of variation is an old idea (see note 4). Yet, evolutionary discourse has proceeded without any rigorous grounding for internalist thinking. In particular, the Haldane-Fisher argument appeared to undermine this kind of theory.

Nevertheless, a rigorous population-genetic grounding for internalist thinking is possible, based on the theory of biases in the introduction process. The logic of this theory has been validated by mathematical and computer modeling. Empirical studies have shown a strong effect of ordinary mutation biases on the changes involved in adaptation, an effect that is expected under the theory but which is not possible under the mutation pressure theory of Haldane and Fisher.

The kinds of variational tendencies covered by the theory include (1) mutation biases such as transition-transversion bias, (2) local asymmetries induced by the properties of genotype-phenotype (GP) maps, and (3) differences in findability and connectivity of phenotypic forms induced by broad features of the architecture of genetic spaces.

When this previously missing causal link is considered in the broader context of contemporary internalist arguments, it provides a way to specify complete chains of causation from internal tendencies of variation to quantifiable tendencies of evolution, via population genetics, rationalizing key themes of internalist thinking: taxon-specific dispositions, directional trends, and intrinsically likely structures.

This doesn’t mean that every internalist claim in the contemporary or historic literature is correct or even meaningful. What it means is that we can specify an alternative to neo-Darwinism that can be used to generate hypotheses that provide a causal grounding for common internalist themes— a complete grounding from internal properties to quantifiable evolutionary tendencies, by way of population genetics. Because the theory is quantitative, it will be possible, ultimately, to make statements that compare the relative importance of internal factors influencing varigenesis with the importance of selection.


  • Alberch P, Gale EA. 1985. A developmental analysis of an evolutionary trend: digital reduction in amphibians. Evolution 39:8-23.
  • Amundson R. 2001. Adaptation, Development, and the Quest for Common Ground. In:  Hecht S, Orzack SH, editors. Adaptation and Optimality. New York: Cambridge University Press. p. 303-334.
  • Arthur W. 2004. Biased Embryos and Evolution. Cambridge: Cambridge University Press.
  • Eimer T. 1898. On Orthogenesis; and The Impotence of Natural Selection in Species-Formation. Chicago: Open Court Publishing Co.
  • Eshel I, Feldman MW. 2001. Optimality and Evolutionary Stability under Short-term and Long-term Selection. In:  Orzack SH, Sober E, editors. Adaptationism and Optimality. Cambridge: Cambridge University Press. p. 161-190.
  • Fox RF. 1993. Review of Stuart Kauffman, The Origins of Order: Self-Organization and Selection in Evolution. Biophysical Journal 65:2698-2699.
  • Gould SJ, Lewontin RC. 1979. The spandrels of San Marco and the Panglossian paradigm: a critique of the adaptationist program. Proc. Royal Soc. London B 205:581-598.
  • Hartl DL, Taubes CH. 1998. Towards a theory of evolutionary adaptation. Genetica 103:525-533.
  • Lynch M. 2007. The frailty of adaptive hypotheses for the origins of organismal complexity. Proc Natl Acad Sci U S A 104 Suppl 1:8597-8604.
  • Maynard Smith J. 1983. Evolution and Development. In:  Goodwin BC, Holder N, Wylie CC, editors. Development and Evolution. New York: Cambridge University Press. p. 33-46.
  • Mayr E. 1994. Response to John Beatty. Biology and Philosophy 9:357-358.
  • McCandlish DM, Stoltzfus A. 2014. Modeling evolution using the probability of fixation: history and implications. Quarterly Review of Biology 89:225-252.
  • Michod RE. 1981. Positive Heuristics in Evolutionary Biology. The British Journal for the Philosophy of Science 32:1-36.
  • Mitchell, P. Coupling of Phosphorylation to Electron and Hydrogen Transfer by a Chemi-Osmotic type of Mechanism. Nature 191, 144–148 (1961).
  • Morgan TH. 1910. The American Society of Naturalists Chance or Purpose in the Origin and Evolution of Adaptation. Science 31:201-210.
  • Popov I. 2009. The problem of constraints on variation, from Darwin to the present. Ludus Vitalis 17:201-220.
  • Provine WB. 1978. The role of mathematical population geneticists in the evolutionary synthesis of the 1930s and 1940s. Stud Hist Biol. 2:167-192.
  • Provine WB. 1971. The Origins of Theoretical Population Genetics. Chicago: University of Chicago Press.
  • Shull AF. 1935. Weismann and Haeckel: One Hundred Years. Science 81:443-451.
  • Sober E. 1984. The Nature of Selection: Evolutionary Theory in Philosophical Focus. Cambridge, Mass.: MIT Press.
  • Stoltzfus A. 2006. Mutation-Biased Adaptation in a Protein NK Model. Mol Biol Evol 23:1852-1862.
  • Stoltzfus A. 2019. Understanding bias in the introduction of variation as an evolutionary cause. In:  Uller T, Laland KN, editors. Evolutionary Causation: Biological and Philosophical Reflections. Cambridge, MA: MIT Press.
  • Xue JZ, Costopoulos A, Guichard F. 2015. A Trait-based framework for mutation bias as a driver of long-term evolutionary trends. Complexity 21:331-345.
  • Yampolsky LY, Stoltzfus A. 2001. Bias in the introduction of variation as an orienting factor in evolution. Evol Dev 3:73-83.
  • Yedid G, Bell G. 2002. Macroevolution simulated with autonomously replicating computer programs. Nature 420:810-812.


1. Classical population genetics does not ignore mutation or treat it solely as a background condition. In particular, classic work pays loads of attention to deleterious mutation pressure. That is, in the case of deleterious mutation pressure, mutation is often treated as a change-making causal process characterized by explicit dynamics. But we are not concerned here with deleterious mutation pressure. We are concerned with the novelty-introducing role of mutation, in situations where this role may lead to changes that are actually incorporated in long-term evolution.

2. We should not be surprised by the lack of generality of the Synthesis position, which was never intended to be general, but was constructed deliberately to exclude certain broad classes of ideas. That is, the SGFT and the ideas of causation that emerged mid-century were constructed deliberately to rationalize a neo-Darwinian view and to make alternatives appear unreasonable or impossible. To some extent, the position of Mayr and his cohort of influencers was less like an ordinary scientific theory (driven by the challenge of explaining empirical patterns) and more like a rhetorical battleship with its guns were aimed squarely at alternatives to neo-Darwinism.

In the contemporary literature, the “Synthesis” remains more of a rhetorical strategy than a scientific theory, but now the focus is purely defensive, aimed at developing a flexible (goal-post-shifting) rhetorical strategy to fight off calls for reform, rather than a strategy to vanquish rivals and establish pre-eminence. That is, contemporary defenses of the Synthesis represent a rhetorical posture of defending the fullness and authority of tradition against the claims of reformers, by anchoring all new developments in tradition, without making any risky scientific claims. The common theme is still the TINA doctrine (There Is No Alternative), but originally this meant that one tradition excluded all the others, whereas now it means that one tradition appropriates all valuable ideas and claims ownership of them on behalf of tradition.

Provine, in his 2001 re-issue of The Origins of Theoretical Population Genetics, said that the Modern Synthesis “came unraveled” in the 1980s. Because Provine construed the Synthesis orthodoxy mainly as a position on population genetics, he is most concerned with the breakdown of the SGFT and the rise of neutralism. That is, Provine is not associating the demise of the Synthesis with the paleontology challenge (1970s to 1980s) or the evo-devo challenge (1980s onward), but instead with a changing understanding of population genetics primarily due to the challenge from molecular evolution.

3. Note that the conception and use of forces is typically deterministic, e.g., Sober (1984) says that “In evolutionary theory, mutation and selection are treated as deterministic forces of evolution” whereas drift is treated stochastically.  Importantly, one may aggregate the effects of the introduction process over infinitely many loci, e.g., all the sites in a genome, and this makes it possible to speak in a technically correct way about a mass-action pressure, but it is a pressure of introduction. Many, many claims in the molecular evolution literature refer to “mutation pressure” but do not make any sense unless we reinterpret them in terms of introduction pressure. But introduction pressure is a different kind of pressure from the classical forces, operating in a different field: it aggregates over sites or loci in a genome rather than over member organisms in a population. Because it is an entirely different kind of pressure, it has different implications (for more explanation, read this). It would be possible to articulate an alternative theory of forces for evolutionary behavior in a discrete space where the steps are origin-fixation steps (Pablo Razeto-Barry has an unpublished manuscript on this).

4. This old idea is sometimes called “orthogenesis,” but the relentless caricatures by traditional authorities have given “orthogenesis” such a pejorative connotation that it is not useful to use. For a review of historic ideas about constraints or channeling of variation untainted by Darwinian prejudices, see Popov (2009).

5. I’m giving myself a pass not to foreground quantitative genetics here because it would be confusing. The theory is fundamentally phenomenological rather than causal, in the sense that it was not built in a bottom-up way from mechanisms or causes, but specified in a top-down way by the constraint of relating certain measurable quantities. So, it is always difficult to fathom QG in a discussion of causes, though clearly the original motivation was tied to a neo-Darwinian view of variation as raw materials, i.e., variation (passive object) as a material cause, not varigenesis (active process) as an agent with dispositional effects. Because raw materials are just raw materials, providing substance only and not form, the only meaningful question to ask about them is some version of “how much do I have?” But if variation is seen as a dynamic process operating in a multidimensional space, then we have lots of questions to ask, or (stated differently) lots of ways to parameterize it.

Apropos, after the multivariate generalization of Lande and Arnold (1983), QG was no longer strictly aligned with neo-Darwinism, but became a formalism with the (initially cryptic) potential to support a more causally oriented theorizing with varigenesis as a dispositional factor in evolution via M, although quantitative geneticists themselves have no love for this idea, and it has only a limited scope because the entire framework has a limited scope. The theory, by original conception, only applies to quantitative traits with abundant infinitesimal variation, and the typical implementations treat dimensional heterogeneity but not directional bias. That is, varigenesis (M) represents a process that generates different amounts of abundant infinitesimal raw material in different dimensions. It can generate more variation along some dimension, but biases in one direction (along a dimension) are usually not considered, and when they are considered, they are not found to be important (except Xue, et al 2015 find support for directional trends in a quantitative character, albeit with a non-standard approach).

What the results of Houle, et al. suggest (at least, to me) is a model with a hidden trait C that undergoes episodic changes affecting wing traits idiosyncratically. Suppose wing shape represented by traits A and B (using these 2 traits to exemplify many wing traits) are variable subject to correlations that reflect invariant aspects of wing development. But suppose that the overall pattern of divergence is driven by fluctuating selection on C, a hidden trait that affects wing shape indirectly, in an idiosyncratic way. Each change in C affects A and B, with the usual AB correlations, but there is no systematic AC or BC correlation.

In a standard model, such individual effects are infinitesimal and, being uncorrelated, have no net effect, being swamped by other infinitesimal effects. So, we suppose instead that C undergoes episodic changes with non-infinitesimal effects on wings. Yet, by assumption, these non-infinitesimal effects (on A and B) follow the usual correlations that reflect invariant aspects of wing development unaffected by C. To sample an episode of evolutionary change under this model is to sample a vector of change in AB space with a direction that tends to align with Gmax for A and B, and a magnitude that depends on the degree of non-infinitesimalness of the wing effects from episodic changes in C.

The following crude and risky experiment could be use to validate some aspects of this model. Define C concretely as the composite trait implicated by a list of (for example) 20 idiosyncratic adaptive effects known in fruit flies, e.g., resistance to DDT, resistance to alcohol, some clinal polymorphism, etc. To the extent that such effects trace to specific mutations, they can be crossed easily into the lab stock to measure the effect on wing shape as M’. To the extent that M’ looks like M for spontaneous mutations, the above scenario is biologically plausible.

The ultimate point here is that quantitative genetics also provides a possible theoretical grounding for internalist thinking, but this grounding is of such limited utility that quantitative geneticists themselves do not seem to invest much promise in internalist explanations, and I’m finding it quite easy to ignore for my purposes here, e.g., it doesn’t provide a way to rebut the Haldane-Fisher argument or justify findability claims. However, I am open to being convinced, by someone who knows better, that QG provides a broader causal grounding for internalism than I have suggested.

6. A strictly verbal theory of biases in the introduction process emerged in an odd place: the macroevolution debate of the 1970s and 1980s. The participants in this debate quickly reached a consensus that no new fundamental mechanisms were needed to account for macroevolution, only a hierarchical expansion of existing mechanisms, i.e., an expansion from the traditional level of a population of individuals, to all the levels in a hierarchy of populations (cells, individuals, species, higher taxa).

However, in the process of this expansion, some of the participants creatively misinterpreted traditional thinking. Vrba and Eldredge (1984) depicted evolution as a dual process of the introduction and reproductive sorting (by selection and drift) of variants, and they emphasized that evolutionary biases could emerge from either introduction or sorting. Based on this formula, they helpfully reinterpreted evo-devo statements (Rachootin and Thompson; Oster and Alberch) to mean that “bias in the introduction of phenotypic variation may be more important to directional phenotypic evolution than sorting by selection.” That is, their elegantly stated verbal theory (1) recognizes, as distinct phases of the evolutionary process, the production or introduction of variation, and the reproductive sorting of variation (by selection and drift), (2) in parallel, distinguishes biases in introduction from biases in sorting as alternative causes of evolutionary bias, and (3) generalizes this theory of dual causation to multiple levels of a hierarchy.

In this way, Vrba and Eldredge (1984) proposed a novel quasi-mutationist theory of evolution as a process of mutation proposes, sorting disposes. However, the theory lacked any model or formalization, so it was not possible to generate quantitative expectations or offer any proofs. Furthermore, participants in the macroevolution debate did not treat this as a radical proposal demanding validation, because it was presented (mistakenly) as merely a restatement of orthodoxy. That is, whereas Maynard Smith et al recognized the conflict between implicit evo-devo theories and classical population genetics, Vrba and Eldredge did not. Their language continues to reverberate in the paleontology literature, but the issues of causation have never been clarified, to my knowledge.

(7) Here and elsewhere I refer to the introduction process in a general way, because it is a generally useful idea that goes beyond mutational origination. Various processes that we do not normally consider as mutation can introduce discrete genetic novelties, e.g., events of lateral transfer, inter-compartmental transfer, recombination, and endosymbiogenesis. Further, large classes of evolutionary processes feature dynamic dependence on events of introduction. In island biogeography, for instance, we can conceptualize a dual process of introduction (a gravid fly is blown to an island) and establishment (the immigrant fly gives rise to a persistent lineage on the island), such that biases in either stage would be effectual. Adaptive dynamics could be seen as a dual proposal-acceptance (introduction-invasion) process, subject to biases in introduction. Cases of cultural evolution such as the evolution of ideas or of language likewise could be treated with origin-fixation dynamics.

In a discrete world, there is always an event that introduces something novel, i.e., an event that makes the step from a frequency of 0 to a frequency of 1/N. However, in the world of modeling, or perhaps in the physical world, there may be cases in which it is useful to define the introduction process as the transient of a continuous value as it departs from 0. Certainly one may foresee, for the case of evolutionary dynamics, conditions under which the dynamics of this departure from 0 are dominated by the contribution of mutation from other alleles even when other processes are operating simultaneously.

8. I’m 80 % sure that my memory is correct on this. I recently asked a friend in Halifax to verify this by photographing the page with the expletive. Alas, the Kellogg library has discarded old volumes of Nature, so the story cannot be confirmed.

9. Traditionalists will object loudly to any suggestion that traditional authorities were wrong about anything, and will attempt to personalize the issue by making it about the perspicacity of Haldane and Fisher. The more adept ones will quote-mine the ancient literature, e.g., in this case, they could quote-mine mutationist leanings from Haldane or that model from Fisher (1930) in which a new mutation is added every generation, and object on this basis that of course Fisher and Haldane understood the importance of new mutations bla bla bla. However, our arguments here are about theories, not people, and particularly they are about the theories that have shaped evolutionary discourse by being actually used to explain things and win arguments and influence people. If Haldane recanted his 1927, 1932, and 1933 argument in a later piece that no one ever read, this is irrelevant both to science and to scientific history. Haldane’s beliefs as beliefs are irrelevant both to science and to understanding historical evolutionary discourse. Perhaps Haldane was secretly a mutationist, a Christian and a capitalist. Who cares? Meanwhile, the Haldane-Fisher argument is a genuine argument used repeatedly in evolutionary discourse to beat down internalist thinking. It is an argument that matters — in a way that can be documented. So long as the canon that defines the historical discourse on evolution includes such sources as Fisher (1930) paper, Huxley (1942) or Haldane (1932), the Haldane-Fisher argument is a part of this discourse. It is not in any sense a strawman or a misrepresentation. It is a documented and readily recognizable thread woven into the 20th century discourse about evolution.