How are social and epistemic structures driving future interactions in scientific research? We analyzed a very large data set of scientific publications (APS dataset) to investigate the social dynamics of collaboration network and the epistemic network of PACS codes co-occurrences network. In particular we show that the epistemic network built through concepts co-occurrences in articles has a non trivial structure characterized by a scale-free degree distribution and a high value of modularity, produced by linking probability mechanisms that depend on local similarity factors. We empirically show that those local epistemic dynamics are both linked to past social and epistemic structures. Moreover we show that the evolution of the social network depends also on epistemic factors and vice versa, indicating that the two networks mutually influence their evolution and are thus co-evolving over time.