Evolution involves interplay between natural selection and developmental constraints. This is seen, for example, when digits are lost from the limbs during evolution. Extant archosaurs (crocodiles and birds) show several instances of digit loss under different selective regimes, and show limbs with one, two, three, four or the ancestral number of five digits. The 'lost' digits sometimes persist for millions of years as developmental vestiges. Here we examine digit loss in the Nile crocodile and five birds, using markers of three successive stages of digit development. In two independent lineages under different selection, wing digit I and all its markers disappear. In contrast, hindlimb digit V persists in all species sampled, both as cartilage, and as Sox9- expressing precartilage domains, 250 million years after the adult digit disappeared. There is therefore a mismatch between evolution of the embryonic and adult phenotypes. All limbs, regardless of digit number, showed similar expression of sonic hedgehog (Shh). Even in the one-fingered emu wing, expression of posterior genes Hoxd11 and Hoxd12 was conserved, whereas expression of anterior genes Gli3 and Alx4 was not. We suggest that the persistence of digit V in the embryo may reflect constraints, particularly the conserved posterior gene networks associated with the zone of polarizing activity (ZPA). The more rapid and complete disappearance of digit I may reflect its ZPA-independent specification, and hence, weaker developmental constraints. Interacting with these constraints are selection pressures for limb functions such as flying and perching. This model may help to explain the diverse patterns of digit loss in tetrapods. Our study may also help to understand how selection on adults leads to changes in development.

Binding selectivity and cross-reactivity within one of the largest and most abundant interaction domain families, the PDZ family, has long been enigmatic. The complete human PDZ domain complement (the PDZome) consists of 267 domains and we applied here a Bayesian selectivity model to predict hundreds of human PDZ domain interactions, using target sequences of 22,997 non-redundant proteins. Subsequent analysis of these binding scores shows that PDZs can be divided into two genome-wide clusters that coincide well with the division between canonical class 1 and 2 PDZs. Within the class 1 PDZs we observed binding overlap at unprecedented levels, mediated by two residues at positions 1 and 5 of the second α-helix of the binding pocket. Eight PDZ domains were subsequently selected for experimental binding studies and to verify the basics of our predictions. Overall, the PDZ domain class 1 cross-reactivity identified here implies that auxiliary mechanisms must be in place to overcome this inherent functional overlap and to minimize cross-selectivity within the living cell. Indeed, when we superimpose PDZ domain binding affinities with gene ontologies, network topology data and the domain position within a PDZ superfamily protein, functional overlap is minimized and PDZ domains position optimally in the binding space. We therefore propose that PDZ domain selectivity is achieved through cellular context rather than inherent binding specificity.