To evaluate the substrate specificity of methyl-directed mismatch repair in Escherichia coli extracts, we have constructed a set of DNA heteroduplexes, each of which contains one of the eight possible single base pair mismatches and a single hemimethylated d(GATC) site. Although all eight mismatches were located at the same position within heteroduplex molecules and were embedded within the same sequence environment, they were not corrected with equal efficiencies in vitro. G-T was corrected most efficiently, with A-C, C-T, A-A, T-T, and G-G being repaired at rates 40-80% of that of the G-T mispair. Correction of each of these six mispairs occurred in a methyl-directed manner in a reaction requiring mutH, mutL, and mutS gene products. C-C and A-G mismatches showed different behavior. C-C was an extremely poor substrate for correction while repair of A-G was anomalous. Although A-G was corrected to A-T by the mutHLS-dependent, methyl-directed pathway, repair of A-G to C-G occurred largely by a pathway that is independent of the methylation state of the heteroduplex and which does not require mutH, mutL, or mutS gene products. Similar results were obtained with a second A-G mismatch in a different sequence environment suggesting that a novel pathway may exist for processing A-G mispairs to C-G base pairs. As judged by DNase I footprint analysis, MutS protein is capable of recognizing each of the eight possible base-base mismatches. Use of this method to estimate the apparent affinity of MutS protein for each of the mispairs revealed a rough correlation between MutS affinity and efficiency of correction by the methyl-directed pathway. However, the A-C mismatch was an exception in this respect indicating that interactions other than mismatch recognition may contribute to the efficiency of repair.