Personalized drug recommendation with pretrained GNNs on a large biomedical knowledge graph

Prescribing drugs to patients is usually performed based on their condition and the expertise of the medical practitioners. However, finding the most suitable drug is a complex task, as doctors might be influenced not only by the patient’s characteristics but also by factors like drug availability, insurance coverage, and even personal preference. With an increase in data availability and quality from vast knowledge bases focusing on individual traits and biological interactions, computational drug recommendation systems emerged. These tools hold the potential to improve the decision-making for drug prescriptions of a single doctor. In this work, we adapted the graph in-context learning framework Prodigy proposed by Huang et al. to perform personalized drug prediction for individual patients. First, patient-specific knowledge graphs were created using EHR data of patients from UK Biobank and the biomedical knowledge graph PrimeKG. Based on unique patient profiles, incorporating drug-, disease-, genomic-, and demographic information, disease-specific drugs are recommended by using patients with the same disease and known prescriptions as support examples. The performance of our proposed model was evaluated for three psychiatric disorders, namely depression, bipolar disorder, and schizophrenia, by assessing how well the drugs connected with the disease are considered as true over false drugs regarding treatment. Furthermore, we compared the results to a state-of-the-art baseline approach based on an RGAT and DistMult encoder-decoder framework. Even though a direct comparison is difficult due to the different prediction properties of the models, the result of the baseline was used as guidance for performance comparison. Overall, our model achieved higher AUC-ROC values than the disease-specific baseline models. Additionally, besides the comparatively high prediction ability, the model is very flexible for new diseases, only requiring a small number of support patients. Since no disease-specific fine-tuning is required, the calculations can also be conducted very quickly. Finally, as a real-world application, we compared the three highest-ranked drugs recommended by our system against the actually prescribed drugs for one random example.