Recently, Beldade et al. (2002) described an interesting experiment in which they succeeded in artificially selecting for butterflies of the species Bicyclus anynana with large anterior eyespots and small posterior ones (and the converse combination), despite the strong positive correlation between these two characters found in natural populations. Although this experimental result is very clear, I have a concern about the way in which the authors connect it to general evolutionary theory. In particular, I question their main conclusion that their result “argues for a dominant role of natural selection rather than internal constraints.”
To generalize the issue, consider any two characters, X and Y, that are positively correlated within a natural population of any species. In all such situations, the structure of the phenotypic variation can be represented by an upward-sloping ellipse rather than a circle in an X/Y plane. Any environment this population finds itself in will render certain actual or potential combinations of X and Y values more fit than others. The topography of fitness over the X/Y plane of all actual and potential combined character values constitutes the adaptive landscape—a concept with a long history extending back to the early twentieth century (Wright 1932). The direction of phenotypic evolutionary change is not determined by either the structure of the variation or the structure of the adaptive landscape alone—rather it is determined by the relationship between the two. It is therefore a result of the interaction between constraint and selection.
Specifically, positive covariation of X and Y will mean that some adaptive peaks are reached in evolution and others not, despite being equidistant from the population centroid. This is because the biased variation overlaps the slopes leading to some peaks (those close to the upward diagonal) more readily than others (those in the orthogonal dimension). As the species concerned spreads through a range of environments with different adaptive landscapes, the magnitude of this effect will vary. In some cases, the highest neighboring fitness peak will be reached whereas in others it will not.
What Beldade et al. (2002) did show is that it is possible to break a relative (Arthur 2001) or quantitative (Dworkin et al. 2001) constraint by imposing one particular (artificial) adaptive landscape that was chosen specifically to maximize the chances of this happening. Given that experimental design, I am not surprised by their result, and I do not dispute it. What I do dispute is whether it is possible to generalize from this result to the relative importance of natural selection and developmental constraint in the wild.
It may be that Beldade et al. (2002) were interested in the causality of the elliptical pattern of covariation as well as its effects on subsequent evolution. But even in relation to this, their general conclusion is unwarranted. Positive correlations between characters can arise because of (a) intrinsic tendencies of the developmental system, in other words constraint; (b) the effects of previous natural selection; or (c) both. In any of these three cases, it will be possible for natural selection to overcome the prevailing pattern and move the population in other directions. All that is required is for the ellipse to have some finite width; that is, there must not be absolute constraint where the ellipse has given way to a line of positive slope. Providing this is the case, then natural selection may be effective. Whether it actually is so brings us back to the shape of the adaptive landscape and the consequent dangers of extrapolating from one such stepwise artificial landscape to the subtler ones that prevail in nature.
Of course, much hinges on what is meant by constraint, and despite valiant attempts at definitions and clarifications by various authors (e.g. Maynard Smith et al. 1985; Schwenk 1995), we have not yet reached a point of complete clarity on this issue. With regard to different kinds of constraint, it is clear that what is true of relative constraint may not be true of its absolute equivalent. For example, natural selection operating on thousands of species of centipedes over hundreds of millions of years has failed to break the apparently absolute constraint (Arthur and Farrow 1999) that there must be an odd number (between 15 and 191) of leg-bearing segments. As well as clear definitions and subcategories, we need to know much more about the molecular and cellular mechanisms through which constraint arises. Although I do not expect Beldade and co-authors to agree with my earlier paragraphs, I hope they may agree with this final one.