Designing a floorplan is an iterative process that includes room layout design, which is crucial in considering user preferences. If a floorplan does not consider customer preferences, it may result in disorderly room usage and accessibility. This paper describes a neural-guided method that incorporates deep learning and optimization techniques to automatically generate realistic room layouts and meet the given bubble diagram constraints. The topology and geometry of the floorplans are decoupled and predicted individually using dual graph neural networks. A hybrid optimization algorithm is employed to enhance the prediction of orthogonal and planar room layouts. This paper contributes to the practice of generating realistic room layouts with consistent topology, showing strong performance in shape and topology similarity metrics on a public floorplan dataset.