Abstract
The advent of genome-scale genetic and genomic studies allows new insight into disease classification. Recently, a shift was made from linking diseases simply based on their shared genes towards systems-level integration of molecular data. Here, we aim to find relationships between diseases based on evidence from fusing all available molecular interaction and ontology data. We propose a multi-level hierarchy of disease classes that significantly overlaps with existing disease classification. In it, we find 14 disease-disease associations currently not present in Disease Ontology and provide evidence for their relationships through comorbidity data and literature curation. Interestingly, even though the number of known human genetic interactions is currently very small, we find they are the most important predictor of a link between diseases. Finally, we show that omission of any one of the included data sources reduces prediction quality, further highlighting the importance in the paradigm shift towards systems-level data fusion.
Introduction
Traditional disease classification has historically relied on anatomical manifestation or clinical symptoms. While the Disease Ontology (DO) serves as a gold standard for this categorization, it faces challenges in incorporating the vast influx of "omics" data—genomics, proteomics, and metabolomics. These data sources are inherently heterogeneous, noisy, and incomplete, making it difficult to connect disparate dots, such as linking a metabolic pathway to a specific cancer.
To address this, researchers developed a "data fusion" framework that integrates 11 diverse molecular and clinical datasets. By looking past surface-level symptoms to underlying systems-level molecular data, the team identified 14 novel disease-disease associations and demonstrated that even small amounts of genetic interaction data are incredibly powerful predictors of disease relationships.
Methodology: Data Fusion via Matrix Tri-Factorization
The core of this approach is a technique called simultaneous penalized matrix tri-factorization. This method treats each data source (e.g., gene-disease links or drug-target interactions) as a large, sparse matrix.
The "fusion" occurs through factor sharing: different matrices share common dimensions. For instance, if one matrix relates genes to drugs and another relates genes to diseases, the model forces them to share the same "gene factor," allowing information to flow across the entire system. The researchers used existing hierarchies, such as the Disease Ontology, as constraints to guide the model. To ensure robustness, consensus clustering was employed, retaining only those associations that appeared consistently across multiple runs.
Data Integration and Systems-Level Synergy
The study integrated 11 distinct data sources covering four primary object types:
- Genes: Protein-protein interactions (BioGRID), gene co-expression, cell signaling (KEGG), and metabolic networks.
- Drugs: Drug-target relationships and drug-drug interactions (DrugBank).
- Ontologies: Hierarchical structures of both the Gene Ontology (GO) and Disease Ontology (DO).
- Associations: Known gene-disease relationships (Mapped GeneRIF) and gene-annotation links.
A key finding was the "synergy" of these sources: omitting any single data source—even noisy ones—reduced the model's accuracy. This validates the systems-level approach, proving that the whole is truly greater than the sum of its parts.
Key Results
- Molecular Reconstruction of Medicine: The model successfully grouped diseases into 108 highly stable clusters based on molecular evidence, largely reconstructing the manually curated Disease Ontology. This proves that clinical categories have deep, measurable roots in molecular systems.
- The Power of Genetic Interactions: Despite being one of the smallest datasets (511 interactions), the removal of genetic interaction data caused the largest drop in predictive quality—a 9.5% decrease in explained variance and a 13.3% increase in error.
- Novel Associations: The team identified 14 disease associations not currently present in the Disease Ontology but supported by literature and comorbidity data. Notable examples include links between gastric lymphoma and crescentic glomerulonephritis, and Cushing’s syndrome and Hodgkin’s lymphoma.
Conclusion and Future Directions
While the model identifies strong associations, it does not inherently distinguish between causation and correlation. For example, a link might stem from a shared genetic mechanism or be a side effect of a common treatment. Future work aims to integrate more granular data, such as Genome-Wide Association Studies (GWAS), to refine these links. The ultimate goal is a data-driven hierarchy of disease that assists in predicting comorbidities and tailoring personalized therapies based on molecular profiles.
Summary updated