Lacking mechanistic disease definitions and corresponding association data hamper progress in network medicine and beyond

Sepideh Sadegh, James Skelton, Elisa Anastasi, Andreas Maier, Klaudia Adamowicz, Anna Cäcilie Möller, Nils M. Kriege, Jaanika Kronberg, Toomas Haller, Tim Kacprowski, Anil Wipat, Jan Baumbach, David B. BlumenthalNature Communications 2023 · DOI

TL;DR

Network medicine aims to move beyond symptom-based disease definitions to understand underlying molecular mechanisms. However, this paper reveals that using large-scale, phenotype-based disease association data for this purpose can introduce significant biases, leading to a "blurred view" of true disease mechanisms. The study emphasizes the need for caution when using such data and suggests combining it with detailed molecular data from well-characterized patient cohorts.

What Problem Does This Paper Address?

The field of network medicine strives to redefine diseases based on their underlying molecular mechanisms (endotypes) rather than observable symptoms (phenotypes). This shift is crucial for developing more targeted and effective treatments. A key approach in network medicine involves building networks from large-scale disease association data (e.g., gene-disease links, comorbidity data) to mine for these underlying mechanisms. However, a critical problem arises: these large-scale datasets often use the very phenotype-based definitions that network medicine aims to overcome. This creates a circular problem where the data used to discover mechanistic subtypes is itself based on imprecise, symptom-driven classifications. The authors investigate the extent to which these inherent biases in disease annotations distort the findings of network medicine studies, particularly those employing a "bird's-eye-view" (BEV) approach.

Methods

The researchers employed a comparative network analysis approach using Graph Edit Distance (GED). GED is a measure of dissimilarity between two graphs. They constructed various networks, or "diseasomes" and "drugomes," based on different types of disease association data (e.g., disease-gene, disease-variant, comorbidity, drug-indication).

To test their hypotheses, they performed two main types of analyses:

  1. Global-scale analysis: They compared entire networks constructed from different data types. The hypothesis was that networks derived from data assumed to contain mechanistic information should be more similar to each other than expected by chance. This was assessed using global GED.
  2. Local-scale analysis: They zoomed in on the immediate neighborhood of individual disease or drug nodes within these networks. The hypothesis was that these local neighborhoods should also be more similar between networks than expected by chance. This was assessed using local GED.

To establish a baseline for comparison and assess statistical significance, they generated 1000 randomized "permuted" networks for each real network. They then compared the GED values of real networks to those of their permuted counterparts. They also conducted analyses using different disease vocabularies (MONDO, UMLS CUI, ICD-10) to assess the impact of granularity.

Datasets

The study utilized several publicly available databases for network construction, including:

  • DisGeNET: For disease-gene and disease-variant associations.
  • OMIM: For disease-gene associations.
  • DrugCentral: For drug-target and drug-indication information.
  • CTD: For drug-indication data.
  • UniProt: For protein-protein interactions.
  • Estonian Biobank: For comorbidity data (using ICD-10 codes).
  • HPO (Human Phenotype Ontology): For disease-symptom associations.

The study did not introduce new benchmark datasets but rather integrated and analyzed existing ones. They specifically highlighted the challenge of mapping diverse disease vocabularies (e.g., OMIM, MeSH, ICD-10) to common target vocabularies (MONDO, UMLS CUI, ICD-10), noting that this mapping process itself can lead to data loss.

Key Insights and Ideas

  • The "Blurred View" Problem: Large-scale disease association data, when annotated with phenotype-based definitions, provides a "blurred" or imprecise view of underlying molecular disease mechanisms. This is a fundamental limitation for network medicine's goal of discovering endotypes.
  • Global vs. Local Similarity: While networks built from different data types show significant similarity on a global scale (supporting the overall validity of BEV network medicine), this similarity breaks down at the local level (individual disease neighborhoods). This suggests that BEV approaches can give a general overview but struggle with detailed mechanistic insights at the individual disease level.
  • Drug Data is More Reliable Locally: Analyses involving drug-disease and drug-drug networks showed better preservation of local neighborhoods compared to disease-disease networks. This suggests that drug-related annotations might be more mechanistically precise or less affected by phenotype-based biases.
  • Variant Data Shows Promise: Disease-variant associations appeared to provide more targeted information locally than disease-gene associations, potentially due to the nature of GWAS studies and the finer granularity of variant data.
  • Vocabulary Matters: The granularity and choice of disease vocabulary significantly impact the results, with coarser vocabularies (like ICD-10) sometimes showing stronger local similarities, possibly because they aggregate diseases that might share some underlying mechanisms, albeit imprecisely.

Key Results

  • Global-scale hypothesis strongly supported: Networks constructed from different data types consistently showed significant similarity on a global scale, corroborating previous findings and the general utility of BEV network medicine.
  • Local-scale hypothesis only partially supported: While some local similarities were observed, a substantial fraction of disease nodes did not show preserved neighborhoods, indicating that the local-scale hypothesis does not hold universally.
  • ICD-10 vocabulary showed more significant local similarities: Comparisons using the ICD-10 vocabulary (a coarser granularity) generally yielded more significant local similarities than those using MONDO or UMLS CUI, suggesting that while imprecise, these broader categories might capture some shared mechanistic signals better than very fine-grained, but potentially sparse, classifications.
  • Drug neighborhoods better preserved: Local analyses showed that drug neighborhoods were more consistently preserved across different network types than disease neighborhoods.

Limitations and Future Work

  • Confounders: The authors acknowledge that other confounders besides mechanistically inadequate disease definitions might influence their results (e.g., off-target effects in drug data, biases in gene association data). However, they stress the stability of their results across different data modalities, suggesting phenotype-based definitions are a primary driver of the observed blurriness.
  • Unsupervised Methods Needed: The paper calls for the development of unsupervised network medicine methods that can de novo stratify patients into mechanistically distinct subgroups without relying on potentially misleading phenotypic annotations.
  • Need for Close-up Approaches: The study strongly advocates for combining BEV approaches with "close-up" analyses that utilize detailed, patient-level molecular and phenotypic data, especially for well-characterized patient cohorts.
  • Future Work: The authors suggest further investigation into specific disease chapters within vocabularies to understand patterns of local blurriness and speculate that diseases where current definitions already align well with true endotypes might be where the local-scale hypothesis holds. They also developed a web tool (GraphSimViz) and a Python package (GraphSimQT) to help researchers explore these findings and assess the reliability of different data sources for their specific research questions.

Summary updated

Abstract

A long-term objective of network medicine is to replace our current, mainly phenotype-based disease definitions by subtypes of health conditions corresponding to distinct pathomechanisms. For this, molecular and health data are modeled as networks and are mined for pathomechanisms. However, many such studies rely on large-scale disease association data where diseases are annotated using the very phenotype-based disease definitions the network medicine field aims to overcome. This raises the question to which extent the biases mechanistically inadequate disease annotations introduce in disease association data distort the results of studies which use such data for pathomechanism mining. We address this question using global- and local-scale analyses of networks constructed from disease association data of various types. Our results indicate that large-scale disease association data should be used with care for pathomechanism mining and that analyses of such data should be accompanied by close-up analyses of molecular data for well-characterized patient cohorts.
BibTeX
@inproceedings{sadegh2023lacking,
  author = {Sepideh Sadegh and James Skelton and Elisa Anastasi and Andreas Maier and Klaudia Adamowicz and Anna Cäcilie Möller and Nils M. Kriege and Jaanika Kronberg and Toomas Haller and Tim Kacprowski and Anil Wipat and Jan Baumbach and David B. Blumenthal},
  title = {{Lacking mechanistic disease definitions and corresponding association data hamper progress in network medicine and beyond}},
  booktitle = {Nature Communications},
  year = {2023},
  doi = {10.1038/s41467-023-37349-4},
}
Figure