During the course of cancer development, a normal cell progresses toward malignancy by acquiring a specific series of mutations. These include mutations that activate otherwise innocuous proto-oncogenes, and others that inactivate recessive tumor suppressor genes. By acquiring these mutations, a cell progressively alters its phenotype, and thereby eludes the various controls that normally prevent malignant growth in an organism. Based on epidemiological data (Renan 1993), and consistent with in vitro experimental data (Hahn et al. 1999), it is estimated that between four and eight rate-limiting mutations occur during the development of most human cancers. But this raises a conundrum. The incidence of cancer should be proportional to the number of rate-limiting events necessary for tumorigenesis, the frequency of these events, and the size of the target cell population for these events. Therefore, given that somatic mutations arise at a frequency of< 6× 10− 6 per locus (Seshadri et al. 1987), an overly simplistic calculation would suggest that even a tumor requiring only four mutations would only arise at a frequency of∼ 1 in 10− 21 cells, a vanishingly low frequency even in an organism composed of∼ 1014 cells, as humans are. Why, then, are the odds of developing cancer during one’s lifetime∼ 1 in 3, and what does this tell us about the mechanisms that operate during tumorigenesis? At least two factors confound the simplicity of the above calculation. The first concerns the frequency of occurrence of the rate-limiting events. In tumors, genes may be epigenetically silenced by methylation as well as inactivated by genetic mutation (Jones and Laird 1999), and, therefore, the mutation rate does not necessarily accurately reflect the overall rate of gene inactivation. Furthermore, nearly all human tumors show an enhanced mutation rate or chromosomal instability (Lengauer et al. 1998). The total frequency of gene inactivation in tumor cells may therefore be up to 1000-fold higher than in normal cells (Seshadri et al. 1987). A second confounding factor concerns the size of the target cell population that undergoes mutation. If a single mutation, either activation of an oncogene or inactivation of a tumor suppressor gene, were to confer a significant selective advantage on an emerging tumor cell, the resulting clonal expansion would vastly increase the cellular target size for each subsequent mutation, and thus increase the overall frequency of tumorigenesis (eg, see Fig. 1B). Is the high incidence of cancer in human populations telling us that this second mechanism also must play a significant role in cancer development in vivo? The appearance of characteristic mutations at each increasingly malignant stage of colon carcinoma progression (Fearon and Vogelstein 1990) suggests that selection for intermediate-stage, premalignant genetic lesions does, in fact, occur. However, the precise details of how this occurs remain unclear. For example, although one might expect that mutations that activate oncogenes would confer a selective advantage on a cell, by driving cell cycle progression and/or by inducing genomic instability, the converse appears, paradoxically, to be true as well: in normal cells, overexpression of oncogenes such as Ras, Myc, and E1a activates an intracellular tumor surveillance mechanism (see below) and results in cell cycle arrest or apoptosis rather than clonal expansion (Serrano et al. 1997; de Stanchina et al. 1998; Zindy et al. 1998). It is also not obvious how inactivating mutations in tumor suppressor genes would confer a selective advantage. Although complete loss of a tumor suppressor gene function could provide a selective …