Linear infrastructures threaten ecosystems and have both direct and indirect effects on ecological habitats and individuals. The fragmentation effect (i.e. splitting or loss of habitat patches) and the barrier effect (i.e. fences, traffic noise avoidance, and roadkill) are widely reported. Wildlife-vehicle collisions are the most visible direct effect and can result in severe human injuries. In Europe, the roe deer population is growing and roe deer roadkills are becoming more common. Roe deer movements depend on landscape features and regional-scale connectivity. Here, we investigate the influence of the landscape network on the location of roe deer roadkill hotspots. In order to localize new potential hotspots along the national roads of Franche-Comté (Eastern France), we first show that roe deer roadkills are not distributed randomly and we identify hotspots. Then, we explain roe deer hotspot locations using a predictive model by combining landscape composition variables, road-related properties, and graph-based connectivity metrics. We test three centrality metrics at three dispersal distances and we assess the relative contribution of the connectivity metrics to the best model. In order to define high-risk sections, we find the probability that reduces the costs of misclassification that the model produces. We validate the model with a new set of roe deer roadkills. Finally, to help the road manager to prioritize mitigation measures, the three 500m road sections with the highest value of probability defined by the model were identified. These sections are located near well-known hotspots.