How epistasis hinders or facilitates movement on fitness landscapes has been a longstanding question of interest. Several high throughput experiments have demonstrated that despite its idiosyncrasy, epistatic effects exhibit global statistical patterns. Recently, Papkou et. al. constructed a fitness landscape for a 9-base region in the folA gene, which encodes for dihydrofolate reductase (DHFR), in E. coli, and demonstrated that despite being highly rugged, the landscape is highly navigable. In this work, using the folA landscape, we ask two questions: (1) How does the nature of epistatic interactions change as a function of the genomic background? (2) How predictable is epistasis within a gene? Our results show that epistasis is 'fluid' - the nature of epistasis exhibited by a pair of mutations is strongly contingent on the genetic background. Mutations exhibit one of two binary 'states': a small fraction of mutations exhibit extremely strong patterns of global epistasis, while most do not. Despite these observations, we observe that the distribution of fitness effects (DFE) of a genotype is highly predictable based on its fitness. These results offer a new perspective on how epistasis operates within a gene, and how it can be predicted.