The role of epistasis – inter-dependent contributions of alleles to fitness – in shaping genetic variation within and between populations is an important question in evolutionary biology with significant implications for our understanding of the factors contributing to phenotypic variation. While epistasis has been shown to play an important role in evolutionary processes such as speciation and adaptive evolution, many aspects of this role remains poorly understood. In particular, there is much debate on whether observing prevalent epistasis in evolution can be taken as evidence for functional epistasis that is relevant to selectable variation. Here, we studied the nature of epistasis in protein evolution, and found a high prevalence of epistatic interactions between amino acid sites in the human genome. We showed that these interactions can help improve accuracy of predicting the impact of genetic variation on the protein structure and function. We also showed that hypothesis-driven search for epistasis in natural populations can detect genomic signatures of epistasis in humans.