Predicting traffic accidents plays an important role in improving transportation efficiency and urban safety. Among other things, the built environment and road network design of a city can affect the incidence of traffic accidents. Existing studies tend to focus on abstract theories, posing challenges for urban designers to apply intuitively. This study proposes a workflow to analyze and predict urban traffic accidents by integrating urban road networks, land use, and building profiles. We extracted data on traffic accident occurrences in San Francisco in 2016–2023 and mapped the coordinates to the city’s land use and road network maps, and developed a graph-born graph prediction model for urban traffic accidents using GAN neural networks. The model was able to produce fairly accurate predictions of traffic accidents in the city of San Francisco. We used the model to analyze the corresponding road safety situations under common urban prototypes and road network patterns from the perspective of urban design (road modeling), and summarized the impacts of common road and land use patterns on traffic safety in the city of San Francisco, as well as the possible ways to improve them. In addition, the model is applied to other cities in the U.S. to validate the model’s migration capability.