Most existing models treat genomic and clinical data in isolation, limiting our ability to personalize care and interpret genetic variation in biologically meaningful ways. Graph-based machine learning offers a powerful framework for modeling the complex, interconnected structure of biological systems.
Key gaps persist in the field:
* Treatment decisions for common, polygenic diseases are often empirical and do not account for genomic profiles.
* A large proportion of genetic variants identified in the population remain classified as variants of uncertain significance (VUS).
* Variant interpretation is not factoring in the broader genomic context of the person in whom the variant is observed.
To address these challenges, we will build a comprehensive biological knowledge graph incorporating curated relationships among genetic variants, genes, proteins, pathways, drugs, and phenotypes. We will integrate UK Biobank participant data into this graph to enable personalized, context-aware analyses. Our objectives are:
1. To personalize medication selection for individuals based on their genomic and phenotypic profile by modelling treatment ranking as a link prediction task between patient nodes and drug nodes in the graph.
2. To develop graph neural network models predict variant pathogenicity factoring in the biological and clinical context of the individual, improving the interpretability of VUS in a personalized framework.
3. To identify subgroups of patients with shared molecular, clinical, or treatment response characteristics through graph-based clustering, uncovering biologically meaningful disease subtypes.
This project will contribute novel methods and graph architecture for integrating structured biological knowledge with individual-level data, helping to advance the goals of precision medicine.