Evaluating knowledge graphs for disease prediction and therapeutic drug recommendation using proteomics data

Introduction: Proteomics is important for understanding disease mechanisms and identifying therapeutic targets by linking proteins to various biological aspects. However, linking proteins with their function and further information (e.g., protein-drug interactions) requires more sophisticated approaches and data structures. Implementing knowledge graphs (KGs) could systematically organize and standardize protein data and further information, enhancing the accessibility and clarity of protein connections. Therefore, KGs potentially make automated decision processes possible such as automated diagnosis or treatment customization for individual patients. Recent KGs, such as the Clinical Knowledge Graph, BioDWH2-KG, BioKG, PharMeBINet, and PrimeKG, have been developed to incorporate proteomics information and connect them with various biological aspects.

Material and Methods: Our objective is to test these five KGs with clinical proteomics data to evaluate their current capabilities and identify limitations. Specifically, we aim to determine if KGs can connect patient proteomics data with corresponding diseases, identify key disease processes based on the patient data, and assist in identifying therapeutic drugs based on dysregulated proteins involved in key disease processes. This evaluation should provide an overview of how these KGs perform in these tasks and assess whether it is possible to automate diagnosis and treatment customization for individual patients, or what additional developments are needed to achieve this. For testing, we utilized diverse patient datasets, including results from a meta-study on inflammatory bowel disease (IBD) biomarkers, proteomics data from pre- and post-drug treatment for IBD, and data from patients with polyneuropathy and platelet disorders.

Results: Our first results indicate that PharMeBINet provides the most comprehensive information for connecting proteins and drugs, resulting in the best recommendations for therapeutic drugs. However, a combined approach using PharMeBINet, Clinical Knowledge Graph, and BioDWH2-KG significantly enhances the quality of insights due to their varied data integration methods and diverse database coverage. Predicting diseases using solely proteomics data faces challenges, primarily due to a bias stemming from better-researched diseases having more connections and a higher chance to connect to the patient data. Consequently, disease prediction for less-studied diseases, which are poorly connected, tends to yield poorer outcomes.

Conclusion: This study will comprehensively evaluate existing KGs with incorporated proteomics data, examining their potential to automate disease diagnosis and drug recommendation processes.