The Extended Phenotype

by Richard Dawkins

  The reason I may sound reductionistic is that I insist on an atomistic view of units of selection, in the sense of the units that actually survive or fail to survive, while being whole-heartedly interactionist when it comes to the development of the phenotypic means by which they survive:

  Of course it is true that the phenotypic effect of a gene is a meaningless concept outside the context of many, or even all, of the other genes in the genome. Yet, however complex and intricate the organism may be, however much we may agree that the organism is a unit of function, I still think it is misleading to call it a unit of selection. Genes may interact, even “blend”, in their effects on embryonic development, as much as you please. But they do not blend when it comes to being passed on to future generations. I am not trying to belittle the importance of the individual phenotype in evolution. I am merely trying to sort out exactly what its role is. It is the all important instrument of replicator preservation: it is not that which is preserved [Dawkins 1978a, p. 69].

  In this book I am using the word ‘vehicle’ for an integrated and coherent ‘instrument of replicator preservation’.

  A vehicle is any unit, discrete enough to seem worth naming, which houses a collection of replicators and which works as a unit for the preservation and propagation of those replicators. I repeat, a vehicle is not a replicator. A replicator’s success is measured by its capacity to survive in the form of copies. A vehicle’s success is measured by its capacity to propagate the replicators that ride inside it. The obvious and archetypal vehicle is the individual organism, but this may not be the only level in the hierarchy of life at which the title is applicable. We can examine as candidate vehicles chromosomes and cells below the organism level, groups and communities above it. At any level, if a vehicle is destroyed, all the replicators inside it will be destroyed. Natural selection will therefore, at least to some extent, favour replicators that cause their vehicles to resist being destroyed. In principle this could apply to groups of organisms as well as to single organisms, for if a group is destroyed all the genes inside it are destroyed too.

  Vehicle survival is only part of the story, however. Replicators that work for the ‘reproduction’ of vehicles at various levels might tend to do better than rival replicators that work merely for vehicle survival. Reproduction at the organism level is familiar enough to need no further discussion. Reproduction at the group level is more problematic. In principle a group may be said to have reproduced if it sends off a ‘propagule’, say a band of young organisms who go out and found a new group. The idea of a nested hierarchy of levels at which selection might take place—vehicle selection in my terms—is emphasized in Wilson’s (1975) chapter on group selection (e.g. his figure 5-1).

  I have previously given reasons for sharing in the general scepticism about ‘group selection’ and selection at other high levels, and nothing in the recent literature tempts me to change my mind. But that is not the point at issue here. The point here is that we must be clear about the difference between those two distinct kinds of conceptual units, replicators and vehicles. I have suggested that the best way of understanding Eldredge and Gould’s theory of ‘species selection’ is in terms of species as replicators. But the majority of models ordinarily called ‘group selection’, including all those reviewed by Wilson (1975), and most of those reviewed by Wade (1978), are implicitly treating groups as vehicles. The end result of the selection discussed is a change in gene frequencies, for example an increase of ‘altruistic genes’ at the expense of ‘selfish genes’. It is still genes that are regarded as the replicators which actually survive (or fail to survive) as a consequence of the (vehicle) selection process.

  As for group selection itself, my prejudice is that it has soaked up more theoretical ingenuity than its biological interest warrants. I am informed by the editor of a leading mathematics journal that he is continually plagued by ingenious papers purporting to have squared the circle. Something about the fact that this has been proved to be impossible is seen as an irresistible challenge by a certain type of intellectual dilettante. Perpetual motion machines have a similar fascination for some amateur inventors. The case of group selection is hardly analogous: it has never been proved to be impossible, and never could be. Nevertheless, I hope I may be forgiven for wondering whether part of group selection’s enduring romantic appeal stems from the authoritative hammering the theory has received ever since Wynne-Edwards (1962) did us the valuable service of bringing it out into the open. Anti-group selectionism has been embraced by the establishment as orthodox, and, as Maynard Smith (1976a) notes, ‘It is in the nature of science that once a position becomes orthodox it should be subjected to criticism …’ This is, no doubt, healthy, but Maynard Smith drily goes on: ‘It does not follow that, because a position is orthodox, it is wrong …’ More generous recent treatments of group selection are given by Gilpin (1975), E. O. Wilson (1975), Wade (1978), Boorman and Levitt (1980), and D. S. Wilson (1980, but see criticism by Grafen 1980).

  I am not going to go again into the debate over group selection versus individual selection. This is because the main purpose of this book is to draw attention to the weaknesses of the whole vehicle concept, whether the vehicle is an individual organism or a group. Since even the staunchest group selectionist would agree that the individual organism is a far more coherent and important ‘unit of selection’, I shall concentrate my attack on the individual organism as my representative vehicle, rather than on the group. The case against the group should be strengthened by default.

  It may seem that I have invented my own concept, the vehicle, as an Aunt Sally to be easily knocked down. This is not so. I simply use the name vehicle to give expression to a concept which is fundamental to the predominant orthodox approach to natural selection. It is admitted that, in some fundamental sense, natural selection consists in the differential survival of genes (or larger genetic replicators). But genes are not naked, they work through bodies (or groups, etc.). Although the ultimate unit of selection may indeed be the genetic replicator, the proximal unit of selection is usually regarded as something larger, usually an individual organism. Thus Mayr (1963) devotes a whole chapter to demonstrating the functional coherence of the whole genome of an individual organism. I shall discuss Mayr’s points in detail in Chapter 13. Ford (1975, p. 185) disdainfully writes off the ‘error’ that ‘the unit of selection is the gene, whereas it is the individual’. Gould (1977b) says:

  Selection simply cannot see genes and pick among them directly. It must use bodies as an intermediary. A gene is a bit of DNA hidden within a cell. Selection views bodies. It favors some bodies because they are stronger, better insulated, earlier in their sexual maturation, fiercer in combat, or more beautiful to behold … If, in favoring a stronger body, selection acted directly upon a gene for strength, then Dawkins might be vindicated. If bodies were unambiguous maps of their genes, then battling bits of DNA would display their colors externally and selection might act upon them directly. But bodies are no such thing … Bodies cannot be atomized into parts, each constructed by an individual gene. Hundreds of genes contribute to the building of most body parts and their action is channeled through a kaleidoscopic series of environmental influences: embryonic and postnatal, internal and external.

  Now if this were really a good argument, it would be an argument against the whole of Mendelian genetics, just as much as against the idea of the gene as the unit of selection. The Lamarckian fanatic H. G. Cannon, indeed, explicitly uses it as such: ‘A living body is not something isolated, neither is it a collection of parts as Darwin envisaged it, like, as I have said before, so many marbles in a box. That is the tragedy of modern genetics. The devotees of the neo-Mendelian hypothesis regard the organism as so many characters controlled by so many genes. Say what they like about polygenes—that is the essence of their fantastic hypothesis’ (Cannon 1959, p. 131).

  Most people would agree that this is not a good argument against Mendelian genetics, and no more i
s it a good argument against treating the gene as the unit of selection. The mistake which both Gould and Cannon make is that they fail to distinguish genetics from embryology. Mendelism is a theory of particulate inheritance, not particulate embryology. The argument of Cannon and Gould is a valid argument against particulate embryology and in favour of blending embryology. I myself give similar arguments elsewhere in this book (e.g. the cake analogy in the section of Chapter 9 called ‘The poverty of preformationism’). Genes do indeed blend, as far as their effects on developing phenotypes are concerned. But, as I have already emphasized sufficiently, they do not blend as they replicate and recombine down the generations. It is this that matters for the geneticist, and it is also this that matters for the student of units of selection.

  Gould goes on:

  So parts are not translated genes, and selection doesn’t even work directly on parts. It accepts or rejects entire organisms because suites of parts, interacting in complex ways, confer advantages. The image of individual genes, plotting the course of their own survival, bears little relationship to developmental genetics as we understand it. Dawkins will need another metaphor: genes caucusing, forming alliances, showing deference for a chance to join a pact, gauging probable environments. But when you amalgamate so many genes and tie them together in hierarchical chains of action mediated by environments, we call the resultant object a body.

  Gould comes closer to the truth here, but the truth is subtler, as I hope to show in Chapter 13. The point was alluded to in the previous chapter. Briefly, the sense in which genes may be said to ‘caucus’ and form ‘alliances’ is the following. Selection favours those genes which succeed in the presence of other genes, which in turn succeed in the presence of them. Therefore mutually compatible sets of genes arise in gene-pools. This is more subtle and more useful than to say that ‘we call the resultant object a body’.

  Of course genes are not directly visible to selection. Obviously they are selected by virtue of their phenotypic effects, and certainly they can only be said to have phenotypic effects in concert with hundreds of other genes. But it is the thesis of this book that we should not be trapped into assuming that those phenotypic effects are best regarded as being neatly wrapped up in discrete bodies (or other discrete vehicles). The doctrine of the extended phenotype is that the phenotypic effect of a gene (genetic replicator) is best seen as an effect upon the world at large, and only incidentally upon the individual organism—or any other vehicle—in which it happens to sit.

  7 Selfish Wasp or Selfish Strategy?

  This is a chapter about practical research methodology. There will be those who accept the thesis of this book at a theoretical level but who will object that, in practice, field workers find it useful to focus their attention on individual advantage. In a theoretical sense, they will say, it is right to see the natural world as a battleground of replicators, but in real research we are obliged to measure and compare the Darwinian fitness of individual organisms. I want to discuss a particular piece of research in detail to show that this is not necessarily the case. Instead of comparing the success of individual organisms, it is often in practice more useful to compare the success of ‘strategies’ (Maynard Smith 1974) or ‘programs’ or ‘subroutines’, averaged across the individuals that use them. Of the many pieces of research that I could have discussed, for instance the work on ‘optimal foraging’ (Pyke, Pulliam & Charnov 1977; Krebs 1978), Parker’s (1978a) dungflies, or any of the examples reviewed by Davies (1982), I choose Brockmann’s study of digger wasps purely because I am very familiar with it (Brockmann, Grafen & Dawkins 1979; Brockmann & Dawkins 1979; Dawkins & Brockmann 1980).

  I shall use the word ‘program’ in exactly the same sense as Maynard Smith uses ‘strategy’. I prefer ‘program’ to ‘strategy’ because experience warns that ‘strategy’ is quite likely to be misunderstood, in at least two different ways (Dawkins 1980). And incidentally, following the Oxford English Dictionary and standard American usage, I prefer ‘program’ to ‘programme’ which appears to be a nineteenth-century affectation imported from the French. A program (or strategy) is a recipe for action, a set of notional instructions that an animal seems to be ‘obeying’, just as a computer obeys its program. A computer programmer writes out his program in a language such as Algol or Fortran, which may look rather like imperative English. The machinery of the computer is so set up that it behaves as if obeying these quasi-English instructions. Before it can run, the program is translated (by computer) into a set of more elementary ‘machine language’ instructions, closer to the hardware and further from easy human comprehension. In a sense it is these machine language instructions that are ‘actually’ obeyed rather than the quasi-English program, although in another sense both are obeyed, and in yet another sense neither is!

  A person watching and analysing the behaviour of a computer whose program had been lost might, in principle, be able to reconstruct the program or its functional equivalent. The last four words are crucial. For his own convenience he will write the reconstructed program in some particular language—Algol, Fortran, a flow chart, some particular rigorous subset of English. But there is no way of knowing in which, if any, of these languages the program was originally written. It may have been directly written in machine language, or ‘hard-wired’ into the machinery of the computer at manufacture. The end result is the same in any case: the computer performs some useful task such as calculating square roots, and a human can usefully treat the computer as if it was ‘obeying’ a set of imperative instructions written out in a language that is convenient for humans to understand. I think that for many purposes such ‘software explanations’ of behaviour mechanisms are just as valid and useful as the more obvious ‘hardware explanations’ favoured by neurophysiologists.

  A biologist looking at an animal is in somewhat the same position as an engineer looking at a computer running a lost program. The animal is behaving in what appears to be an organized, purposeful way, as if it was obeying a program, an orderly sequence of imperative instructions. The animal’s program has not actually been lost, for it never was written out. Rather, natural selection cobbled together the equivalent of a hard-wired machine code program, by favouring mutations that altered successive generations of nervous systems to behave (and to learn to change their behaviour) in appropriate ways. Appropriate means, in this case, appropriate for the survival and propagation of the genes concerned. Nevertheless, although no program was ever written down, just as in the case of the computer running a program which has been lost, it is convenient for us to think of the animal as ‘obeying’ a program ‘written’ in some easily understood language such as English. One of the things we can then do is to imagine alternative programs or subroutines which might ‘compete’ with each other for ‘computer time’ in the nervous systems of the population. Though we must treat the analogy with circumspection, as I shall show, we can usefully imagine natural selection as acting directly on a pool of alternative programs or subroutines, and treat individual organisms as temporary executors and propagators of these alternative programs.

  In a particular model of animal fighting, for example, Maynard Smith (1972, p. 19) postulated five alternative ‘strategies’ (programs):

  1 Fight conventionally; retreat if opponent proves to be stronger or if opponent escalates.

  2 Fight at escalated level. Retreat only if injured.

  3 Start conventionally. Escalate only if opponent escalates.

  4 Start conventionally. Escalate only if opponent continues to fight conventionally.

  5 Fight at escalated level. Retreat before getting hurt if opponent does likewise.

  For the purpose of computer simulation it was necessary to define these five ‘strategies’ more rigorously, but for human understanding the simple imperative English notation is preferable. The important point for this chapter is that the five strategies were thought of as if they (rather than individual animals) were competing entities in their own right. Ru
les were set up in the computer simulation for the ‘reproduction’ of successful strategies (presumably individuals adopting successful strategies reproduced and passed on genetic tendencies to adopt those same strategies, but the details of this were ignored). The question asked was about strategy success, not individual success.

  A further important point is that Maynard Smith was seeking the ‘best’ strategy in only a special sense. In fact he was seeking an ‘evolutionarily stable strategy’ or ESS. The ESS has been rigorously defined (Maynard Smith 1974), but it can be crudely encapsulated as a strategy that is successful when competing with copies of itself. This may seem an odd property to single out, but the rationale is really very powerful. If a program or strategy is successful, this means that copies of it will tend to become more numerous in the population of programs and will ultimately become almost universal. It will therefore come to be surrounded by copies of itself. If it is to remain universal, therefore, it must be successful when competing against copies of itself, successful compared with rare different strategies that might arise by mutation or invasion. A program that was not evolutionarily stable in this sense would not last long in the world, and would therefore not present itself for our explanation.