How do the fitness effects of several mutations combine? Despite its simplicity, this question is central to the understanding of multilocus evolution. Epistasis (the interaction between alleles at different loci), especially epistasis for fitness traits such as reproduction and survival, influences evolutionary predictions 1,2 ''almost whenever multilocus genetics matters'' 3 . Yet very few models 4,5 have sought to predict epistasis, and none has been empirically tested. Here we show that the distribution of epistasis can be predicted from the distribution of single mutation effects, based on a simple fitness landscape model 6 . We show that this prediction closely matches the empirical measures of epistasis that have been obtained for Escherichia coli 7 and the RNA virus vesicular stomatitis virus 8 . Our results suggest that a simple fitness landscape model may be sufficient to quantitatively capture the complex nature of gene interactions. This model may offer a simple and widely applicable alternative to complex metabolic network models, in particular for making evolutionary predictions.Recent technical improvements in genetics have allowed the measurement of epistatic interactions in a very precise way. Large amounts of empirical evidence stemming from the study of development, metabolic networks 5 and quantitative traits analyses 9 have accumulated to show that epistasis is a widespread feature of genetic systems. However, despite many examples of epistatic interactions between particular pairs of loci, relatively little is known about the overall distribution of epistasis among random sets of mutations scattered across the genome. A few studies have sought to measure this distribution directly in model systems such as Escherichia coli 7 , RNA viruses 8,10,11 and Saccharomyces cerevisiae 12 . These studies have shown that the variance of epistatic interactions is large compared with their mean, which is always relatively close to zero 10 .Unfortunately, theoretical developments have not gone hand-inhand with those empirical advances, and in particular, no theory is yet available to explain, predict or generalize those observations. Fitness epistasis among mutations at enzymatic loci has been modeled with metabolic control theory 4 or flux balance analysis 5 . The former assumes idealized metabolic pathways and specific metabolism-fitness relationships, whereas the latter models a much more precise and complete metabolic network based on genomic data from model