Recent advances in genome research have accelerated the process of locating candidate genes and the variable sites within them and have simplified the task of genotype measurement. The development of statistical and computational strategies to utilize information on hundreds -- soon thousands -- of variable loci to investigate the relationships between genome variation and phenotypic variation has not kept pace, particularly for quantitative traits that do not follow simple Mendelian patterns of inheritance. We present here the combinatorial partitioning method (CPM) that examines multiple genes, each containing multiple variable loci, to identify partitions of multilocus genotypes that predict interindividual variation in quantitative trait levels. We illustrate this method with an application to plasma triglyceride levels collected on 188 males, ages 20--60 yr, ascertained without regard to health status, from Rochester, Minnesota. Genotype information included measurements at 18 diallelic loci in six coronary heart disease--candidate susceptibility gene regions: APOA1--C3--A4, APOB, APOE, LDLR, LPL, and PON1. To illustrate the CPM, we evaluated all possible partitions of two-locus genotypes into two to nine partitions (approximately 10(6) evaluations). We found that many combinations of loci are involved in sets of genotypic partitions that predict triglyceride variability and that the most predictive sets show nonadditivity. These results suggest that traditional methods of building multilocus models that rely on statistically significant marginal, single-locus effects, may fail to identify combinations of loci that best predict trait variability. The CPM offers a strategy for exploring the high-dimensional genotype state space so as to predict the quantitative trait variation in the population at large that does not require the conditioning of the analysis on a prespecified genetic model.