Digital transformation in the construction industry has led to the introduction and use of heterogeneous software systems. Unified interfaces are essential to enable their use. These require a high degree of standardisation and have little flexibility in mapping data, so interfaces are often a bottleneck for data. Parallel modelling results in asynchronous data processing and often redundant data that must be manually reconciled on an ongoing basis. This unconnected data in different software systems leads to the accumulation of heterogeneous, distributed, and inaccessible data sources and knowledge silos. The presented approach aims at developing a single source of truth in structural design based on a graph database using labelled property graphs (LPG). In the context of this work, the suitability of graphs for the holistic, digital representation of structural building design data is investigated by integrating structural design mindsets and processes. In addition to an overarching methodology showing the future integration of a graph database into existing structural design data exchange structures, a data structure is developed for mapping structural design data into a LPG. A prototypical implementation demonstrates that a graph-based digital representation of structural design data is suitable for data management and exchange of basic structural data, and can be extended to process-oriented data. The focus of graph databases on relationships between data corresponds to the high relevance of connections between components within the design. The developed approach integrates the previous way of working with partial models and positional planning, and also enables bi-directional data exchange between BIM and structural analysis throughout the process. This allows for holistic and interconnected data management without loss of context and data exchange on a unified database. This form of linked data also enables the use of data-driven machine learning approaches, particularly graph-based approaches such as graph neural networks.