A gene library was constructed coding for all possible variants of two amino acids (101, 102) in a solvent-exposed surface return loop (alpha E-beta D) of Bacillus stearothermophilus L-lactate dehydrogenase (bsLDH). All but one of 38 enzyme variants examined were thermally stable and had native-like hydrodynamic properties. In this sample, there was no bias detected in either the DNA or amino acid sequences encoded. We argue that the alpha E-beta D surface loop sequence is unimportant for protein folding or stability and can be fully varied to select enzymes with new substrate specificities. The selection of NAD-dependent dehydrogenases with specificity for: malate, phenyllactate, hydroxyisocaproate and 4-phenyl-2-hydroxy-butanoate from two bsLDH libraries is described. This required a highly discriminatory screen for 2-hydroxy acid dehydrogenase activity to select enzymes which, in the absence of the natural allosteric activator fructose-1,6-bisphosphate (FBP), maintained high temperature stability and catalytic activity without substrate inhibition. In general the amino acid residues at positions 101 and 102 which determined substrate specificity were as expected from hydrophobic and ionic complementarity to the substrate. For example, a bsLDH variant with Asn101Va1102 is as efficient with phenylpyruvate as is the wild-type enzyme (Asn101Gln102) with pyruvate. Using molecular modelling, the valine at position 102 can be fitted into the active site without significant structural distortion caused by the aromatic side-chain of the substrate. Similarly, nine out of ten malate dehydrogenases (MDHs) selected had an arginine residue at position 102 to complement the negatively charged carboxyl group in oxaloacetate. One, Arg101Arg102 (Kcat/Kmoxaloacetate = 1.6 x 10(6) M-1 S-1) is 25% more active than the previous best synthetic MDH. There were surprises: present understanding would not have predicted the oxaloacetate transforming activity of Ser101Leu102 or the phenylpyruvate activity of Pro101Lys102. The former is about one-third as efficient as the best malate dehydrogenase selected, whilst the latter had about one-seventh of the best phenylpyruvate dehydrogenase activity.