Restriction-modification (RM) systems are believed to have evolved to protect cells from foreign DNA. However, this hypothesis may not be sufficient to explain the diversity and specificity in sequence recognition, as well as other properties, of these systems. We report that the EcoRI restriction endonuclease-modification methylase (rm) gene pair stabilizes plasmids that carry it and that this stabilization is blocked by an RM of the same sequence specificity (EcoRI or its isoschizomer, Rsr I) but not by an RM of a different specificity (PaeR7I) on another plasmid. The PaeR7I rm likewise stabilizes plasmids, unless an rm gene pair with identical sequence specificity is present. Our analysis supports the following model for stabilization and incompatibility: the descendants of cells that have lost an rm gene pair expose the recognition sites in their chromosomes to lethal attack by any remaining restriction enzymes unless modification by another RM system of the same specificity protects these sites. Competition for specific sequences among these selfish genes may have generated the great diversity and specificity in sequence recognition among RM systems. Such altruistic suicide strategies, similar to those found in virusinfected cells, may have allowed selfish RM systems to spread by effectively competing with other selfish genes.A type II restriction endonuclease makes a double-strand break within or near a specific recognition sequence in duplex DNA. A cognate modification enzyme methylates the recognition sequence to protect it from the cleavage (1, 2). It is widely accepted that the evolution and maintenance of restriction-modification (RM) systems have been driven by the protection from foreign DNA that they afford to cells. The RM systems do protect cells from infection with some viruses by cleaving their DNA (for example, see ref.3) and are likely to be responsible both for the evolution of antirestriction mechanisms and for the paucity of some restriction sites in certain viruses and plasmids (4).Recent experimental and theoretical analyses (5, 6), however, seem to us to bring into question the efficacy of virusmediated selection for RM systems. Defense by RM systems is short-lived because invading viral DNA will occasionally escape restriction and will become modified, thus affording protection from restriction to itself and its descendants (5, 6). Bacteria will more likely develop other, longer-lasting means of resistance to viruses, such as alterations in the receptor required for infection (5, 6). Although RM systems can provide bacteria with advantage when they are invading new habitats full of phages, it is not clear whether such colonization selection is realistic under natural conditions (5, 6).It is also unclear whether the above "cellular defense" hypothesis can account for the following properties of type II RM systems (1, 2). (i) Their individual high specificity and collective wide diversity in the sequence recognition. (ii) The tight linkage of cognate restriction and modification ...