In this study, the combined effects of geometrical distribution and geomechanical deformation of fracture networks on fluid flow through fractured geological media are investigated numerically. We consider a finite-sized model domain in which the geometry of fracture systems follows a power-law length scaling. The geomechanical response of the fractured rock is simulated using a hybrid finite-discrete element model, which can capture the deformation of intact rocks, the interaction of matrix blocks, the displacement of discrete fractures and the propagation of new cracks. Under far-field stress loading, the locally variable stress distribution in the fractured rock leads to a stress-dependent variable aperture field controlled by compression-induced closure and shear-induced dilatancy of rough fractures. The equivalent permeability of the deformed fractured rock is calculated by solving for the fracture-matrix flow considering the cubic relationship between fracture aperture and flow rate at each local fracture segment. We report that the geometrical connectivity of fracture networks plays a critical role in the hydromechanical processes in fractured rocks. A well-connected fracture system under a high stress ratio condition exhibits intense frictional sliding and large fracture dilation/opening, leading to greater rock mass permeability. However, a disconnected fracture network accommodates much less fracture shearing and opening, and has much lower bulk permeability. We further propose an analytical solution for the relationship between the equivalent permeability of fractured rocks and the connectivity metric (i.e. percolation parameter) of fracture networks, which yields an excellent match to the numerical results. We infer that fluid flow through a well-connected system is governed by traversing channels (forming an “in parallel” architecture) and thus equivalent permeability is sensitive to stress loading (due to stress-dependent fracture permeability), whilst fluid flow through a disconnected system is more ruled by matrix (linking isolated clusters “in series”) and has much less stress dependency.