that contribute to adaptive traits and that will be useful in an agronomic context. These advantages include the following: (i) segregating variation is not required to identify genes of interest; (ii) far fewer plant samples are needed than for LD mapping, with only tens (<100) of samples (Teshima et al., 2006) often sufficing as opposed to hundreds or thousands (Long and Langley, 1999); (iii) like LD mapping, bottom-up approaches can be applied to species that reproduce slowly and lack genetic tools; and (iv) they allow inferences about demographic history, providing historical insights into the process of domestication. We should note, of course, that bottom-up and top-down approaches are not mutually exclusive; for example, bottom-up approaches in maize are also being used to identify candidate genes for LD mapping (Yu and Buckler, 2006).
Although they have advantages, bottom-up approaches are also not a panacea, for at least four reasons. First, their success will vary among species, depending on levels and distribution of genetic diversity. For example, initial surveys of genetic diversity in sorghum have failed to identify selected genes (Hamblin et al., 2006). This failure is in part a limitation of the study system, because sorghum has low genetic diversity, but it may also reflect inefficient sample design. Simulation studies suggest that these methods should be quite powerful with moderate (<100) sample sizes, even with diversity levels as low as those found in sorghum (Teshima et al., 2006, and M. Przeworski, personal communication), but empirical studies relied on a sample of 17 domesticated individuals and only one wild plant (Hamblin et al., 2006). Second, genes identified as selected may have been targets of selection or may be linked to a target of selection (through “hitchhiking”). For example, selection on the rice gene waxy appears to have affected patterns of sequence diversity in 29 additional genes. This lack of resolution is, however, a shortcoming shared with QTL and LD methods, because in all cases it is difficult to differentiate between a target (or “causal”) marker and linkage effects (Weigel and Nordborg, 2005). In fact, when the genomic locations of genes are available, the expected chromosomal resolution of bottom-up approaches is at worst similar to QTL and LD mapping. Third, like top-down approaches, bottom-up approaches may not be feasible for all crops. The limitation here is not generation time (as in QTL studies), but rather levels of genetic diversity, polyploidy, and population structure. Polyploidy makes population genetic analysis difficult, requiring careful separation of homeologs and their independent evolutionary histories. As with LD mapping, unrecognized population structure can be problematic for population genetic analyses, producing patterns that can be mistaken for selection. Finally, bottom-up approaches share a major limitation with both association and QTL mapping. All three methods identify candidate genes or regions, but verification requires additional functional characterization (Weigel and Nordborg, 2005). It is worth noting that in many cases