The Structure of Evolutionary Theory

by Stephen Jay Gould

  There are many further questions on the meaning and limits of clade selection. One issue is whether the populations that bear the gene pools need be in ecological competition with each other. I believe that this is [Page 625] not required, any more than individuals within a population need inter­act ecologically to be subject to individual selection. The reproductive success or failure of a soil arthropod, with an expected lifetime dispersal of a few meters, will hardly influence prospects for a conspecific a hun­dred meters away. But the descendants of these two individuals might compete, and genes passed on by one may ultimately prevail over those passed on by the other. Selective elimination of one and survival of the other a hundred meters away is individual selection as long as the two arthropods can be assigned to the same population and their genes to the same gene pool. ... In the same way, two gene pools in allopatry can be subject to natural selection if, as must always be true, their descendants might be alternatives for representation in the biota . . . The ultimate prize for which all clades are in competition is representation in the biota.

  The internal incoherence of gene selectionism

  I regard the heyday of gene selectionism as an unusual episode in the history of science — for I am convinced that the theory's central argument is logically incoherent, whatever the attraction (and partial validity) of several tenets, and despite the value of a mental exercise that tries to reconceptualize all na­ture from a gene's point of view. Close textual analysis* of this theory's lead­ing documents reveals persistent internal problems, explicitly recognized by authors and invariably met by arguments so flawed in construction that even the defenders seem embarrassed, or at least well aware of the glaring insuf­ficiency.

  I am not alone in noting this peculiar situation, and in calling for some seri­ous consideration by historians. Wilson and Sober (1994, p. 590) write: “The situation is so extraordinary that historians of science should study it in de­tail: a giant edifice is built on the foundation of genes as replicators, and therefore as the 'fundamental' unit of selection, which seems to obliterate the concept of groups as organisms. In truth, however, the replicator concept cannot even account for the organismic properties of individuals. Almost as [Page 626] an afterthought, the vehicle concept is tacked onto the edifice to reflect the harmonious organization of individuals, but it is not extended to the level of groups.”

  The central problem lies as deep as our definition of the key concept of “cause” in science. Aristotle proposed a broad concept of causality divided into four aspects, which he called material, efficient, formal and final (or, roughly, stuff, action, plan and purpose — that is, the bricks, the mason, the blueprint and the function, in the standard “parable of the house,” used for more than two millennia to explicate Aristotle's concept). As many historians have noted, modern science may virtually be defined by a revision of this broad view, and a restriction of “cause,” as a concept and definition, to the aspect that Aristotle called “efficient.” (The word “efficient” derives from the Latin facere, to make or to do. Efficient causes are actual movers and shakers, the agents that apply the forces. Aristotle's term does not engage the modern English meaning of doing something well, as opposed to doing something at all.)

  The Cartesian or Newtonian worldview, the basis of modern science, banned final cause for physical objects (while retaining the concept of pur­pose for biological adaptation, so long as mechanical causes, rather than con­scious external agencies, could be identified — a problem solved by natural se­lection in the 19th century). As for Aristotle's material and formal causes, these notions retained their relevance, but lost their status as “causes” under a mechanical worldview that restricted causal status to active agents. The material and formal causes of a house continue to matter: brick or sticks fash­ion different kinds of buildings, while the bricks just remain in a pile, absent a plan for construction. But we no longer refer to these aspects of building as “causes.” Material and formal attributes have become background condi­tions or operational constraints in the logic and terminology of modern sci­ence.

  I present this apparent digression because the chief error of gene selectionism lies in a failed attempt to depict genes as efficient causes in ordinary natural selection — and the chief “textual mark” of failure can be located in tortuous and clearly discomforting (even to the authors!) arguments ad­vanced by all leading gene selectionists in a valiant struggle to “get through” this impediment. For no matter how an author might choose to honor genes as basic units, as carriers of heredity to the next generation, as faithful repli­cators, or whatever, one cannot deny a fundamental fact of nature: in or­dinary, garden-variety natural selection — Darwin's observational basis and legacy — organisms, and not genes, operate as the “things out there” that live and die, reproduce or fail to propagate, in the interaction with environ­ments that we call “natural selection.” Organisms act as Aristotle's efficient causes — the actors and doers — in the standard form of Darwin's great and universal game.

  Gene selectionists know this, of course — so they must then struggle to construct an argument for saying that, even though organisms do the explicit work, genes may somehow still be construed as primary “units of selection,” [Page 627] or causal agents in the Darwinian process. This misguided search arises from a legitimate intuition — that genes are vitally important in evolution, and clearly central to the process of natural selection — followed by the false infer­ence that genes should therefore be designated as primary causes. Needless to say, no biologist wishes to deny the centrality or importance of genes, just as this intuition holds. But genes simply cannot operate as efficient causes in Darwin's process of organismic selection. Genes, as carriers of continuity to the next generation, may be designated as material causes in Aristotle's aban­doned terminology. But we no longer refer to the material aspects of natural processes as “causes.” Organisms “struggle” as agents or efficient causes; their “reward” may be measured by greater representation of their genes, or material legacies, in future generations. Genes represent the product, not the agent — the stuff of continuity, not the cause of throughput.

  The standard gambit of gene selectionists, in the light of this recognized problem, invokes two arguments, both indefensible.

  Attempts to assign agency to genes by denying emergent properties to organisms. Once one admits, as all gene selectionists must and do, that genes propagate via selection on organisms as interactors, how then can one possibly ascribe direct causal agency to genes rather than to bodies? Only one logical exit from this conundrum exists: the assertion that each gene stands as an optimal product in its own place, and that bodies im­pose no consequences upon individual genes beyond providing a home for joint action. If such a view could be defended, then bodies would become pas­sive aggregates of genes — mere packaging — and selection on a body could then be read as a convenient shorthand summary for selection on all resident genes, considered individually.

  But such a reductionistic view can only apply if genes build bodies with­out nonlinear or nonadditive interactions in developmental architecture. Any nonlinearity precludes the causal decomposition of a body into genes con­sidered individually — for bodies then become, in the old adage, “more than the sum of their parts.” In technical parlance, nonlinearity leads to “emer­gent” properties and fitness at the organismic level — and when selection works upon such emergent features, then causal reduction to individual genes and their independent summations becomes logically impossible. I trust that the empirical resolution of this issue will not strike anyone as controver­sial, for we all understand that organisms are stuffed full of emergent fea­tures — an old intuition stunningly affirmed by the first fruits of mapping the human genome (see the full issues of Science and Nature in February 2001 and my own initial reaction for general audiences in Gould, 2001). What else is developmental biology but the attempt to elucidate such nonlinearities? The error of gene selectionists does not
lie in their stubborn assertion of pure additivity in the face of such knowledge, but rather in their concep­tual failure to recognize that this noncontroversial nonlinearity destroys their theory.

  Dawkins admits the apparent problem (1976, p. 40): “But now we seem to have a paradox. If building a baby is such an intricate venture, and if every [Page 628] gene needs several thousands of fellow genes to complete its task, how can we reconcile this with my picture of indivisible genes, springing like immortal chamois from body to body down the ages: the free, untrammeled, and self-seeking agents of life?”

  Dawkins attempts a lame resolution by invoking the quintessentially Oxbridge metaphor of rowing, with the nine men (eight oarsmen and a cox) as genes, and the boat as a body. Of innumerable candidate rowers, we put to­gether the best boat “by random shuffling of the candidates for each posi­tion” — and then running large numbers of trials until the finest combination emerges. Of course the rowers must cooperate in a joint task, but we generate no nonlinearities because localized optimality prevails, and the winning boat ends up with the best possible oarsman in each place. Dawkins then transfers this image back to biology and asserts his view of selection as piecemeal opti­mization — so that each locus (each seat in the boat) eventually houses the best candidate: “Many a good gene gets into bad company, and finds itself sharing a body with a lethal gene, which kills the body off in childhood. Then the good gene is destroyed along with the rest. But this is only one body, and replicas of the same good gene live on in other bodies which lack the lethal gene . . . Many [good genes] perish through other forms of ill luck, say when their body is struck by lightning. But by definition luck, good and bad, strikes at random, and a gene which is consistently on the losing side is not unlucky; it is a bad gene” (1976, p. 41).

  Such a notion of individualized genetic optimality must be rejected as empirically false; but even if true, this concept still wouldn't support the re­quired claim for nonexistence of emergent organismic features. Even Daw­kins admits (in the quotation just above) that selection can only optimize “phenotypic consequences” (1982, p. 237) — and if phenotypes arise (as they do) by complex nonadditivity among genetic effects, then the genes in your body cannot maintain the essential property of independence represented by Dawkins's metaphor of optimal goats, hopping happily and separately across the generations.

  In any case, this false view of organisms as additive consequences of individually optimized genes underlies the familiar metaphorical language developed by Dawkins over the years: “I am treating a mother as a machine programmed to do everything in its power to propagate copies of the genes which reside inside it” (1976, p. 132). Or “A monkey is a machine which pre­serves genes up trees; a fish is a machine which preserves genes in the water; there is even a small worm which preserves genes in German beer mats. DNA works in mysterious ways” (1976, p. 22). These colorful images misstate ac­tual pathways of causality. Organisms work in wondrous ways, and they op­erate via emergent properties that invalidate Dawkins's concept of genes as primary agents.

  The ceteris paribus dodge. When the logic of an argument re­quires that the empirical world operate in a certain manner, and nature then refuses to cooperate, unwavering supporters often try to maintain their advo­cacy by employing the tactic of conjectural “as if.” That is, you admit the failure of [Page 629] complex nature to meet your theoretical needs, but then claim that you will simplify the actual circumstances “as if” the system under study operated in the required way. The classical “as if” argument goes by its Latin title of ceteris paribus, or “all other things being equal.” Ceteris paribus imposes additivity upon a system truly made of complexly interacting parts. You iso­late one factor and state that you will analyze its independent effects by hold­ing all other factors constant.

  Ceteris paribus ranks among the oldest of scholarly devices, an indispens­able tactic for any student of complex systems. I am certainly not trying to mount a general assault upon this venerable and valuable mode of exempli­fication. Two common circumstances define the legitimate domain of ceteris paribus: (1) as a heuristic or exploratory device for approaching systems of such complexity that you don't even know how to think about influences of particular parts, unless you can hypothetically assign all others to a theoreti­cal background of constancy; and (2) as a powerful experimental tool when you can actually hold other factors constant and perturb the system by vary­ing your studied factor alone.

  But ceteris paribus becomes an illegitimate dodge, an invalid prop to make a potentially false argument unbeatable by definition, in systems dominated by nonadditivity — that is, where the very act of holding all other factors con­stant may make your favored factor work in a manner contrary to its actual operation in a real world of interaction. If A conquers B only when the two entities share a field alone, but usually loses to B when C also dwells on the field, and if real fields invariably include C, then we cannot crown A as abso­lutely superior to B on the basis of a single and artificial ceteris paribus trial that excluded C from action and consideration.

  The use of ceteris paribus to support gene selectionism constitutes a simi­lar denial of a known reality. This tactic represents a fallback position after acknowledging the impossibility of asserting a genuine claim for nonaddi­tivity in the translation of genes to organisms. In other words, you admit that massive nonlinearity actually exists, but then state that, for purposes of dis­cussion, you will counterfactually impose ceteris paribus so that genes can be equated with linear effects. For example, Dawkins explicitly invokes the key phrase (in English rather than Latin) in defending his requisite (but fal­lacious) notion that genes may be identified as operating “for” particular parts of phenotypes, thus creating the impression that organisms may be treated as additive aggregates rather than entities defined by nonlinear inter­action.

  For purposes of argument it will be necessary to speculate about genes “for” doing all sorts of improbable things. If I speak, for example, of a hypothetical gene “for saving companions from drowning,” and you find such a concept incredible, . . . recall that we are not talking about the gene as the sole antecedent cause of all the complex muscular con­tractions, sensory integrations, and even conscious decisions, which are involved in saving somebody from drowning. We are saying nothing [Page 630] about the question of whether learning, experience, or environmental influences enter into the development of the behavior. All you have to concede is that it is possible for a single gene, other things being equal and lots of other essential genes and environmental factors being present, to make a body more likely to save somebody from drowning than its allele would (1976, p. 66).

  In another passage (1976, p. 39), Dawkins unwittingly surrenders this neces­sary tactic by admitting that we dare not discuss the interactive web of em­bryonic development, lest we be unable to speak of genes “for” particular as­pects of organismal phenotypes:

  Everybody knows that wheat plants grow bigger in the presence of ni­trate than in its absence. But nobody would be so foolish as to claim that, on its own, nitrate can make a wheat plant. Seed, soil, sun, water, and various minerals are obviously all necessary as well. But if all these other factors are held constant, and even if they are allowed to vary within limits, addition of nitrate will make the wheat plants grow bigger. So it is with single genes in the development of an embryo. An interlocking web of relationships controls embryonic development so complex that we had best not contemplate it.

  As a striking demonstration that ceteris paribus cannot rescue gene selectionism from logical paradoxes and violations of ordinary linguistic usage, Dawkins (1982, p. 164) addresses the problem of how to treat a genetic dele­tion favored by natural selection at the organismic level, if genes represent the fundamental units of selection, and if we must be able to treat each genetic item as a Darwinian individual with a distinct and independent history. If “gene language” must prevail, and if we need to specify the selective value of such a deletion
, what can we call the loss but “a replicating absence”! Should we not, at this point, admit instead that organisms are the relevant causal agents in this case, and that organisms have achieved a selective benefit by the alternate but orthodox genetic route of deletion rather than substitution? Some humans have done well with “plenty of nothing,” but I don't think we should root our ontology in taxonomies for various kinds and forms of faith­fully propagating absences.

  Any organism that happened to experience a random deletion of part of its selfish DNA would, by definition, be a mutant organism. The deletion itself would be a mutation, and it would be favored by natural selection to the extent that organisms possessing it benefited from it, presumably because they did not suffer the economic wastage of space, materials, and time that selfish DNA brings. Mutant organisms would, other things being equal [ceteris paribus again!], reproduce at a higher rate than the loaded down “wild type” individuals, and the deletion would conse­quently become more common in the gene pool. Here we are recognizing that the deletion itself, the absence of the selfish DNA, is itself a replicat­ing entity (a replicating absence!), which can be favored by selection. [Page 631]

  All major proponents of gene selectionism have unintentionally illustrated the theory's incoherence by trying to “cash out” their system, and failing at the crucial point of assigning causal agency in natural selection. For, however these proponents may talk about genes as primary agents or units of selec­tion, they cannot deny that nature's Darwinian action generally unfolds be­tween discrete organisms and their environments. These authors therefore ac­knowledge this basic fact and then tend to lapse into verbal obfuscation on the gene's behalf. I have already noted a prime example in Williams's claim, quoted previously in another context, that organismic selection should be re­garded not “as a level of selection in addition to that of the gene, but as the primary mechanism of selection at the genic level.” But what does this state­ment mean? Williams recognizes organismic selection as the “primary mecha­nism” by which genes pass differentially from one generation to the next. But primary mechanisms are efficient causes in any standard analysis of the logic of science. Williams (1992, p. 38) presents an accurate epitome of selection in the following passage: he states that selection must always operate on interactors (and he knows, as the previous quotation shows, that organisms usually constitute the relevant interactors in cases that he wishes to describe as genic selection); he also recognizes that information must pass to future generations by faithful heredity, and he seems to acknowledge that such bi­ased passage defines the result, not the cause, of selection. Yet he fails to take the final step of acknowledging that these statements debar gene selectionism as the mechanism of evolution. “Natural selection must always act on physi­cal entities (interactors) that vary in aptitude for reproduction, either because they differ in the machinery of reproduction or in that of survival and re­source capture on which reproduction depends. It is also necessary that there be what Darwin called 'the strong principle of inheritance,' so that events in the material domain can influence the codical record. Offspring must tend to resemble their own parents more than those of other offspring. Whenever these conditions are found there will be natural selection.”