gms | German Medical Science

Introduction: Systems biology employs computer-aided modeling of biological processes to recognize systemic relationships and thus to model and understand complex systems on a highly detailed level. Systems medicine transfers established systems biology approaches to the medical domain. This enables, for example, the modeling of molecular processes in the body or to understand disease progression under different therapeutic conditions. In clinical practice, this has the potential to capture complex interactions and multimorbidities that have so far been ignored, and to enable personalized treatments through data-driven solutions. However, due to the regulations and requirements imposed on medical research, systems medicine mostly occurs independently of clinical practice and bypasses practice-oriented medical informatics.

Various use cases have demonstrated the general applicability of the systems biology approach for clinical problem settings. Additionally, extensive databases with open-source, reproducible virtual experiments already provide a far-reaching knowledge base that has not been exhausted for everyday clinical use, e.g. [1].

Concept: To tap into the enormous potential of systems medicine in everyday medicine, computer-aided methods are required to efficiently find insilico experiments of interest and connect them to patient data. The aims of our work are to develop methods that i) allow us to identify insilico experiments in a structured and quality-assured manner, and ii) allow these models to be linked to patient data, enabling personalized forecasts and new forms of therapy. The framework integrates databases of simulation models in everyday clinical practice and tests models in previously unknown clinical problems.

Methods: We will explore the potential of graph neural networks [2] to convert the graph structures of systems medicine models [3] into specialized vector embeddings [4] that enable an effective search for models, sub-models, and their overlaps. These vector embeddings may also allow interfacing with different query types, such as free text or patient profiles. Existing data will be applied to the queried models, and their calibration will be evaluated. Exceptional model mining can help to efficiently find models that are particularly well-calibrated concerning individual model attributes, resulting in models that only partially cover the current clinical question, or contain parts for which data are not available. Finally, machine learning and deep learning methods will be used to combine partial models (e.g., by embedding them in modular neural networks) and to compensate for missing data and irrelevant partial models (e.g., with the help of transfer and contrastive learning, or generative neural nets). One-shot learning approaches will then be explored, which allow the resulting models to be personalized through individual patient histories. This pipeline will be evaluated on data from data integration centers [5].

Conclusion: This project shows the potential of systems medicine models in everyday medicine. It is another important step towards the use of more precise models of biological relationships in complex clinical problem settings, making patient treatment more effective.

Author’s contributions: MB, RH and DW developed the concept. RH performed implementation tests on the knowledge graph. MB designed the conceptual framework for deep integration. All authors wrote the manuscript.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.