Building a knowledge graph to enable precision medicine

Payal Chandak, Kexin Huang, Marinka Žitnik

Abstract

Developing personalized diagnostic strategies and targeted treatments requires a deep understanding of disease biology and the ability to dissect the relationship between molecular and genetic factors and their phenotypic consequences. However, such knowledge is fragmented across publications, non-standardized repositories, and evolving ontologies describing various scales of biological organization between genotypes and clinical phenotypes. Here, we present PrimeKG, a multimodal knowledge graph for precision medicine analyses. PrimeKG integrates 20 high-quality resources to describe 17,080 diseases with 4,050,249 relationships representing ten major biological scales, including disease-associated protein perturbations, biological processes and pathways, anatomical and phenotypic scales, and the entire range of approved drugs with their therapeutic action, considerably expanding previous efforts in disease-rooted knowledge graphs. PrimeKG contains an abundance of 'indications', 'contradictions', and 'off-label use' drug-disease edges that lack in other knowledge graphs and can support AI analyses of how drugs affect disease-associated networks. We supplement PrimeKG's graph structure with language descriptions of clinical guidelines to enable multimodal analyses and provide instructions for continual updates of PrimeKG as new data become available.

Introduction

The quest for precision medicine—personalized diagnostics and targeted treatments—requires a holistic understanding of disease biology. However, biomedical knowledge is currently a "fragmented landscape." Information regarding gene mutations, clinical symptoms, and drug-protein interactions is scattered across thousands of publications, non-standardized databases, and competing ontologies. This fragmentation creates a major bottleneck for researchers attempting to link molecular causes to clinical effects at scale. Existing knowledge graphs often suffer from limited scale, a lack of clinical depth (such as missing contraindications), and entity inconsistency across different databases.

Methods: Constructing PrimeKG

To address these challenges, we developed PrimeKG (Precision Medicine Knowledge Graph), a massive, multimodal resource that unifies fragmented data across ten biological scales. The construction of PrimeKG involved a systematic pipeline to integrate 20 high-quality primary resources, ranging from genomics and proteomics to clinical phenotypes and drug databases.

The systematic pipeline for constructing PrimeKG from 20 primary data resources, including node identification and relationship harmonization.

Key methodological innovations include:

  • Clinical Entity Resolution: To resolve naming inconsistencies (e.g., a disease named differently in genomic vs. clinical databases), we utilized ClinicalBERT. This language model helped group synonymous concepts into 17,080 clinically meaningful disease nodes, collapsing over 22,000 raw concepts.
  • Multimodal Integration: Unlike traditional graphs that only contain nodes and edges, PrimeKG attaches natural language descriptions from sources like the Mayo Clinic and Orphanet to nodes. This enables researchers to leverage both graph-based AI and Large Language Models (LLMs).
  • Relationship Enrichment: We specifically curated "indications," "contraindications," and "off-label use" edges, providing a more nuanced view of drug-disease relationships than standard "treats" labels.

The PrimeKG Knowledge Graph

PrimeKG contains 129,375 nodes and 4,050,249 edges, representing a significant expansion over previous benchmarks. It covers ten major biological scales, including disease-associated protein perturbations, biological processes, pathways, and anatomical regions.

PrimeKG schema illustrating the ten biological scales and a multimodal example of relationships between Autism and the drug Risperidone.

As shown in the figure above, the graph's richness allows for complex queries. One can trace the relationships between a disease (Autism), its associated genes (IL6, OXT), and therapeutic drugs (Risperidone), including specific clinical contexts like contraindications or synergistic interactions. The authors ensured the graph was well-connected by extracting the "giant component," which retained 99.998% of the original edges.

Key Insights and Results

The Power of "Negative" Data

By including over 61,000 contraindication edges, PrimeKG allows AI models to learn not just what a drug should do, but what it must not do. This is a critical requirement for safe clinical applications and distinguishes PrimeKG from graphs that only include positive "treats" labels.

Rare Disease Authority

PrimeKG is a premier resource for studying "orphan" conditions, covering 90.8% of the rare diseases listed in Orphanet. This extensive coverage enables research into diseases that are often neglected due to data scarcity in smaller datasets.

Validation via Drug Repurposing

To test PrimeKG's utility, we analyzed 11 drugs recently approved by the FDA (after the graph's data cutoff). In 8 out of 10 cases, the drug and its new target disease were significantly "closer" in the PrimeKG network than would be expected by random chance. This demonstrates the graph's predictive power for identifying new therapeutic uses for existing drugs.

Conclusion and Future Directions

PrimeKG bridges the "clinic-to-code" gap by providing a mathematically grounded way to unify biological ontologies with clinical practice. While maintaining such a large-scale multimodal graph requires ongoing effort, it provides a foundational resource for the next generation of AI-driven precision medicine. Future work will focus on using PrimeKG in tandem with Graph Neural Networks (GNNs) to identify new biomarkers, characterize disease processes, and predict drug-drug interactions tailored to a patient's unique molecular profile.

Summary updated

Abstract

Developing personalized diagnostic strategies and targeted treatments requires a deep understanding of disease biology and the ability to dissect the relationship between molecular and genetic factors and their phenotypic consequences. However, such knowledge is fragmented across publications, non-standardized repositories, and evolving ontologies describing various scales of biological organization between genotypes and clinical phenotypes. Here, we present PrimeKG, a multimodal knowledge graph for precision medicine analyses. PrimeKG integrates 20 high-quality resources to describe 17,080 diseases with 4,050,249 relationships representing ten major biological scales, including disease-associated protein perturbations, biological processes and pathways, anatomical and phenotypic scales, and the entire range of approved drugs with their therapeutic action, considerably expanding previous efforts in disease-rooted knowledge graphs. PrimeKG contains an abundance of 'indications', 'contradictions', and 'off-label use' drug-disease edges that lack in other knowledge graphs and can support AI analyses of how drugs affect disease-associated networks. We supplement PrimeKG's graph structure with language descriptions of clinical guidelines to enable multimodal analyses and provide instructions for continual updates of PrimeKG as new data become available.
BibTeX
@inproceedings{chandak2023building,
  author = {Payal Chandak and Kexin Huang and Marinka Žitnik},
  title = {{Building a knowledge graph to enable precision medicine}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41597-023-01960-3},
}
Figure