The development of seismic fragility functions for buildings typically relies on simplified modelling methods and the use of indirect engineering demand parameters for the determination of collapse. Through optimisation of the computational analysis cost, and the incorporation of statistically distributed model properties, this paper demonstrates the potential of non-linear finite element models with soil-structure-interaction (SSI) and explicit progressive collapse prediction as a viable alternative. This paper presents an overview of the method developed and its application to an unreinforced masonry (URM) building in the Groningen region of the Netherlands, where an understanding of the risk arising from induced seismicity is required. Latin hypercube sampling is used to generate batches of 300 realisations of an LS-DYNA response history analysis, each selecting from a set of 100 hazard-consistent ground motions. The extent to which the building model can be varied to account for uncertainty in the modelling and real variability within the building is discussed, including properties of the URM-specific material model, failure characteristics of structural connections, and the spatial variation of these aspects across the model extents. The method allows for regression analysis on the extent of collapse directly in the calculation of the fragility function. Key drivers for fragility of the building are identified from the analyses: mostly parameters describing inner leaf masonry material properties. One goal of the work is to better understand the dependency of the fragility on characteristics of the building stock, such that they can be taken into account in the regional risk assessment.