Experimental and simulation studies show that small monomeric proteins fold in one kinetic step, which entails overcoming the free-energy barrier between the unfolded and the native protein through a transition state. Two models of transition state formation have been proposed: a 'nonspecific' one in which it depends on the formation of a sufficient number of native-like contacts regardless of what amino acids are involved, and a 'specific' one, in which it depends on formation of a specific subset of the native structure (a folding nucleus). The latter requires that some amino acids form most of their contacts in the transition state, whereas others only do so on reaching the native conformation. If so, mutations affecting the stability of the transition state nucleus should have a greater effect on the folding kinetics than mutations elsewhere, and the residues involved should be evolutionarily conserved. Lattice-model simulations and experiments suggest that such mutations exist. Here we present a method for determining the folding nucleus of a protein with known structure with two-state folding kinetics. This method is based on the alignment of many sequences designed to fold into the native conformation of a protein to identify the positions where amino acids are most conserved in designed sequences. The method is applied to chymotrypsin inhibitor 2 (CI2), a protein whose transition state has been previously studied by protein engineering. The involvement of residues in folding nucleus of CI2 is clearly correlated with their conservation in design, and the residues forming the nucleus are highly conserved in 23 natural sequences homologous to CI2.