Learning to take it personally: Precision drug repurposing through patient-specific loss on knowledge graphs using Biobank data

Precision medicine requires drug repurposing methods that adapt to individual patient profiles while working within regulatory frameworks. Existing approaches apply uniform models to all patients, only using individual factors as inputs or filters. Our framework instead integrates patient-specific profiles into the learning algorithm through a customized loss function. We combine standard link prediction with UK Biobank data—integrating polygenic risk scores, biomarker expressions, and medical history. Evaluated on a biomedical knowledge graph connecting 61,000+ entities through 1.2+ million relations, our approach improves drug repurposing quality with AUPRC improvements ranging from 1.3 × to 5.4 × across patients. Case studies on Alzheimer’s Disease patients reveal drug candidates with stronger AD evidence and patient-specific mechanisms. Our loss function identifies influential diseases and biomarkers for each patient, enhancing interpretability while providing biologically relevant recommendations tailored to individual profiles. This approach represents a fundamental shift from treating personalization as data preprocessing to embedding it within the learning objective itself. • Algorithm-level personalization. We embed patient context directly into the learning objective via a patient-specific loss that blends standard link prediction with terms guided by polygenic risk scores (PRS) and protein-biomarker deviations, optimizing for an individual rather than the population. • Clinically grounded signals. We integrate UK Biobank–derived PRS, protein biomarker levels, and diagnosis history to tailor drug–disease scores to each patient’s biology and clinical context, anchoring personalization in routinely collectable, real-world data. • Preserved generalization with better rankings. We maintain foundation-model link-prediction quality while substantially improving patient-specific drug repurposing performance (e.g., AUPRC improvements ranging from 1.3 × to 5.4 × across patients), demonstrating effectiveness without sacrificing global metrics. • Interpretability at the patient level. We learn sparse, patient-level weights over diseases and biomarkers that reveal which comorbidities and dysregulated proteins drive recommendations, supporting transparent, clinician-facing interpretation.