How network medicine reveals surprising connections between diseases even with our incomplete biological map
Imagine trying to complete a massive jigsaw puzzle with most of the pieces missing. This is precisely the challenge scientists face when trying to understand human disease through the lens of our molecular interactions.
Our bodies operate through an incredibly complex network of biological interactions—proteins binding to proteins, genes regulating other genes, and molecules communicating in an intricate dance that keeps us healthy. When this network malfunctions, disease occurs. But how do we begin to understand these breakdowns when we can only see a fraction of the complete picture?
Did you know? Our current interactome covers less than 20% of all potential protein interactions, yet scientists can still predict disease relationships from this incomplete map 1 .
Enter a revolutionary approach: network medicine. By mapping diseases onto the human interactome—the grand map of all molecular interactions in our cells—researchers are discovering that seemingly unrelated diseases often have hidden molecular relationships. Even with our currently incomplete interactome, scientists can now predict how diseases are connected, why they often occur together, and what they might reveal about each other's treatments 1 .
This article explores how researchers are filling in the blanks of our biological puzzle to uncover surprising relationships between diseases.
The incomplete interactome, rather than being a barrier to discovery, has become a powerful lens through which to view human disease relationships.
A framework that views diseases as perturbations of cellular networks
Think of the human interactome as the ultimate social network for your cells—but instead of friends sharing photos, it's proteins, genes, and other molecules interacting to keep you alive. This network consists of approximately 141,296 documented physical interactions between 13,460 proteins, though this represents just a fraction of the complete map 1 .
Like a social network, the interactome has hubs (highly connected proteins), clusters (groups of proteins that interact frequently), and distant connections between less familiar partners.
Interactive visualization of disease modules in the human interactome
At the heart of this research lies the disease module hypothesis—the idea that proteins associated with the same disease tend to cluster in the same neighborhood of the interactome 1 8 . Even when we can't see all the connections, these disease modules form functional units that, when disrupted, lead to specific health conditions.
The fascinating implication is that diseases with overlapping network modules likely share biological mechanisms, even if they appear very different on the surface.
You might wonder how we can draw meaningful conclusions from such an incomplete map. The secret lies in sophisticated mathematical approaches that help researchers see beyond the gaps. Through percolation theory—which studies how connections spread through networks—scientists have determined that despite about 80% of disease proteins appearing disconnected from their main module in current maps, these isolated proteins still tend to cluster in the network vicinity of their disease module 1 .
This pattern holds profound implications: diseases whose modules overlap in the interactome tend to share common biological pathways, similar symptoms, and higher likelihood of occurring together in the same patients 1 .
Current interactome maps cover only a fraction of all possible molecular interactions, yet still reveal meaningful disease relationships.
In a groundbreaking 2015 study published in Science, researchers set out to systematically test whether the network locations of disease modules could reveal meaningful relationships between different diseases 1 . Their approach was both ingenious and methodical, tackling the incompleteness problem head-on.
Compiled the most comprehensive human interactome map available with 141,296 interactions between 13,460 proteins.
Gathered information on 299 different diseases, each with at least 20 known associated genes.
Applied percolation theory to account for the incomplete nature of the interactome.
The core of their methodology revolved around a clever measurement called the network-based separation score (sAB). This metric compared the average shortest distance between proteins within the same disease to the average distance between proteins of different diseases 1 .
The formula was elegant in its simplicity:
Where ⟨dAB⟩ represents the average shortest distance between proteins of disease A and disease B, and ⟨dAA⟩ and ⟨dBB⟩ represent the average shortest distance between proteins within the same disease.
A negative separation score indicated overlapping disease modules, while a positive score suggested topologically distinct diseases.
| Separation Score Range | GO Term Similarity | Symptom Similarity | Comorbidity Risk |
|---|---|---|---|
| Overlapping (sAB < 0) | 10-100x higher | ~10x higher | RR ≥ 10 |
| Separated (sAB > 0) | Baseline to lower | Lower than expected | RR ≈ 1 |
| Disease Category | % Genes in Observable Module | Minimum Genes Needed |
|---|---|---|
| All 299 diseases | ~20% | 25 genes |
| Multiple sclerosis | 16% (11 of 69 genes) | 69 genes |
| Diseases with <25 genes | Too fragmented | 25+ genes required |
The research demonstrated that we don't need a complete interactome to extract meaningful insights. The mathematical models showed that only diseases with at least 25 associated genes would form observable modules in the current interactome—a threshold that helped researchers focus on relationships they could reliably detect 1 .
| Research Tool | Function | Examples/Sources |
|---|---|---|
| Interactome Maps | Blueprint of molecular interactions | 141,296 physical interactions between 13,460 proteins 1 |
| Disease Gene Databases | Catalog genes associated with specific diseases | Online Mendelian Inheritance in Man (OMIM), GWAS databases 1 2 |
| Ontology Resources | Standardized biological terms and relationships | Gene Ontology (GO) annotations 1 |
| Pathway Networks | Contextual functional relationships between biological processes | 1,014 non-redundant human pathways |
| AI and Machine Learning | Find patterns and interactions beyond human analytical capacity | Diamond, survivalFM, graph neural networks 3 5 8 |
This toolkit continues to evolve rapidly. Recent advances in machine learning are helping to overcome the limitations of incomplete data. Tools like Diamond help researchers identify which genetic interactions are most likely to be genuine, providing false discovery rates that help prioritize laboratory experiments 3 .
Another approach called survivalFM uses factorization machines—a concept borrowed from recommendation systems—to model how multiple risk factors interact to influence disease development over time 5 .
Perhaps most promisingly, graph representation learning methods are now using artificial intelligence to automatically embed biological networks into compact vector spaces where algebraic operations can reveal functional relationships 8 .
These AI systems can be thought of as "differentiable engines" that implement longstanding principles of systems biology, potentially discovering relationships that would escape human notice 8 .
The incomplete interactome, rather than being a barrier to discovery, has become a powerful lens through which to view human disease.
By acknowledging the gaps in our knowledge while developing sophisticated methods to work within these constraints, scientists have uncovered a fundamental truth: location matters in the cellular universe.
The implications extend far beyond academic curiosity. Understanding disease-disease relationships can help explain why certain conditions cluster in patients, suggest new therapeutic uses for existing drugs, and potentially identify at-risk populations before symptoms appear.
Future Direction: As research progresses, the map will become more complete, but the network approach will continue to reveal the hidden connections that underlie human health and disease.
The greatest promise may lie in what remains to be discovered. As one researcher noted, "AI models are powerful in building this genotype-to-phenotype mapping, by capturing subtle patterns. Once we have this model, it's not the end of the story. It's just the beginning of the story" 3 .
The incomplete interactome has given us a glimpse of the complex connections between diseases, and each new piece we add promises to reveal more surprises in the intricate network of human health.
Ongoing effort to document molecular interactions
Proposed and validated in early 2000s
Introduced in 2015 Science paper 1
Current frontier in network medicine