Selection Cannot See Next Century: The Standing of Evolvability as an Adaptation — Epoche C2
Some lineages appear better equipped than others to produce new form: the vertebrate forelimb was rebuilt into a wing three separate times, in pterosaurs, birds and bats, and some bacterial lineages acquire antibiotic resistance far faster than close relatives do. The natural interpretation is that such lineages possess something — evolvability, the capacity to generate heritable variation that is sometimes useful — and that natural selection built it, because a lineage that evolves faster outcompetes one that does not. This essay sets a quantitative objection against that interpretation, works through the one case where the objection does not apply, examines the experimental system usually offered as decisive, and asks what survives. 1. The proposal, and the quantities that make it more than a metaphor The claim can be assessed only if evolvability can be measured, so the first task is to say what the measurable parts are and what each of them means. Two are standard, one for short timescales and one for long. On the short timescale, the response of a set of traits to one generation of selection is given by the multivariate breeder's equation, $$\Delta\bar{\mathbf{z}} = \mathbf{G}\boldsymbol{\beta},$$ where $\Delta\bar{\mathbf{z}}$ is the change in the vector of trait means over one generation, $\mathbf{G}$ is the matrix of additive genetic variances and covariances among the traits, estimated from the resemblance among relatives, and $\boldsymbol{\beta}$ is the selection gradient — the vector of partial regression coefficients of relative fitness on the traits, which is what makes it a measure of direct selection on each trait holding the others constant. Lande and Arnold's contribution in 1983 was precisely to show that this partial-regression quantity is the one that enters the response equation, so that the observed change decomposes into an ecological part, $\boldsymbol{\beta}$, and a genetic part, $\mathbf{G}$. In one dimension the equation reduces to the familiar $R = h^2 S$, since $\mathbf{G}$ becomes the additive variance $V_A$, and $\boldsymbol{\beta}$ becomes the selection differential $S$ divided by the phenotypic variance $V_P$, giving $\Delta\bar{z} = V_A S/V_P = h^2 S$. Written this way, the geometry does the work. Because the response is $\boldsymbol{\beta}$ passed through $\mathbf{G}$, a direction in trait space along which $\mathbf{G}$ has a small eigenvalue responds hardly at all, however strongly it is selected; a direction with a large eigenvalue responds readily. Short-term evolvability is therefore not a scalar but an orientation: it is the shape of $\mathbf{G}$, and the same population can be highly evolvable in one direction and effectively frozen in another. On the long timescale, $\mathbf{G}$ is itself transient, since the standing variation it summarises is consumed by selection and drift and must be replenished. The replenishment rate is the mutational variance $V_M$, the additive genetic variance introduced by new mutation each generation. Houle, Morikawa and Lynch assembled the available estimates in 1996 and found that the ratio $V_M/V_E$, with $V_E$ the environmental variance — a quantity conventionally called the mutational heritability — has a median of order $10^{-3}$ per generation across a wide span of traits and taxa. Their own argument is that this standardisation is poor for comparison, since $V_E$ varies enormously with the trait and the measurement protocol, and that mean-standardised mutational coefficients of variation should be preferred; the $10^{-3}$ figure is used here as an order of magnitude, not a biological constant. Taken at face value it says that in the absence of selection and drift, mutation alone accumulates a variance comparable to $V_E$ in about $V_E/V_M = 10^{3}$ generations. Against this background Kirschner and Gerhart argued in 1998 that organisms are built so as to raise these quantities. Their proposal is architectural: a set of core processes conserved across enormous phylogenetic distances, joined to one another by weak, tolerant linkages, so that a change in a regulatory input produces a coherent alteration in output rather than a lethal mess; together with exploratory processes, such as the growth of neurites or microtubules, which are stabilised by whatever targets they encounter and so adjust to a changed body plan without a correlated change of their own. Modularity and robustness of this kind make the genotype-to-phenotype map more forgiving, and variation that is more often viable makes a lineage more evolvable. So far this is description, and it is well supported. The claim under examination is the stronger one: that these features exist because they confer evolvability. 2. The objection: the payoff arrives after the selection event The reason the stronger claim is hard to sustain is a matter of timing, and it needs stating precisely rather than as a slogan about foresight. Natural selection is a difference in reproductive success among individuals alive at a given moment. A feature that improves the variation available to descendants many generations hence pays its bearer nothing at that moment. For such a feature to be favoured, one of exactly three things must hold: it carries a concurrent advantage of its own — in which case evolvability is a by-product, and the adaptive explanation is about the concurrent advantage, not about evolvability; it rides to fixation alongside a beneficial mutation that it helped to produce, remaining physically associated with that mutation long enough to be dragged along; the differential is realised at the level of whole lineages, through differential rates of origination and extinction. Route (i) is conceded rather than contested: if robustness is favoured because misfolded proteins are costly now, that is a complete explanation and evolvability plays no part in it. The remaining two are the substantive ones, and they are best taken in reverse order, because route (iii) is th