Prediction of subduction earthquakes mostly relies on interplate coupling models that provide patterns interpreted within the framework of rate-and-state friction laws. However, this framework has been challenged by recent observations, indicating that rheological and geometrical complexities must be considered to fully understand megathrust mechanics. Here, we explore an alternative explanation: we investigate whether the strongly and weakly coupled patches could be related to the distribution of deformation along the plate interface, associated with either basal erosion and/or underplating. Since underplating and basal erosion both impact forearc morphology, the location of such distributed deformation along the plate interface can be inferred from a simple mechanical analysis of the topography. We first present results along the Chilean subduction margin, where the transition from erosive to accretionary margin from North to South is successfully captured. We show that basal erosion occurs preferentially at 15 km depth, while underplating is found at 35 ± 10 km and 60 ± 5 km depths. We also demonstrate that recent major earthquakes along the accretionary part of the margin were limited at their down-dip end by underplating, whereas along the erosive portion of the margin, earthquakes were surrounded by both basal erosion and underplating. We then present results from various subduction zones (Ecuador, Japan, Nankai, Hikurangi) and discuss the major differences between erosive and accretionary margins in terms of plate interface deformation distribution, depth, and frictional properties. These differences are analysed using thermo-mechanical simulations. We show once again that for several recent earthquakes, a heterogeneous plate interface limited their propagation. We therefore propose that long-lived plate interface deformation undergoes aseismic processes. Large earthquakes then propagate along well-localized and smoothed rate-weakening fault planes, which are bounded by elongate zones of underplating, and are stopped by regions of heterogeneously distributed deformation. Consequently, slip deficit patterns and earthquake segmentation could reflect the along-dip and along-strike distribution of plate interface deformation.