Computational Drug Repurposing with Multiscale Interactomes
Drug repurposing offers a rapid, cost-effective route to new therapies by identifying novel uses for existing compounds. When paired with multiscale interactome analysis, it becomes possible to explore the complex molecular relationships between drugs, targets, pathways, and diseases at a systems level.
This workflow outlines how to construct, analyze, and exploit a heterogeneous biomedical network to identify repurposing candidates, combining graph-based learning with experimental prioritization.
1. Building the Interactome
The foundation of the approach is a heterogeneous graph that integrates multiple layers of biological knowledge. Nodes can represent drugs, proteins, pathways, or diseases, while edges encode interactions—drug–target binding, protein–protein interactions, pathway memberships, and disease associations.
Sources like DrugBank provide curated drug–protein relationships, while protein–protein interaction maps can be drawn from high-confidence databases. For disease biology, integrate experimentally validated or literature-reported links, such as host–pathogen PPIs for infectious diseases or α-synuclein interactors in Parkinson’s disease.
2. Anchoring the Disease Context
Once the network is assembled, the specific disease of interest is added as a node and connected to known associated proteins or pathways. This grounding ensures that the graph captures both molecular interactions and the functional context of the disease.
3. Learning from Network Structure
Graph embedding techniques transform the raw network into a form suitable for machine learning. A Node2Vec pre-processing step captures both local neighborhoods and broader network context. These embeddings are then refined through a Graph Convolutional Network (GCN), trained with a diffusion-based loss function that clusters nodes according to their network proximity to the disease node.
The result is a set of optimized vector representations for every entity in the network—drugs, proteins, and pathways—enabling quantitative similarity searches.
4. Prioritizing Candidates
With embeddings in hand, candidate ranking is straightforward: compute cosine similarity between the disease node and other nodes. This yields:
- Drug proximity scores – for direct repurposing candidates.
- Protein proximity scores – highlighting potential new therapeutic targets.
5. Multiple Selection Strategies
Three complementary selection approaches can be applied:
- Target-centric: Choose drugs directly most similar to the disease node.
- Protein-centric: Identify top-ranked proteins, then retrieve predicted binders from platforms like PolypharmDB.
- Polypharmacology-focused: Prioritize drugs predicted to act on multiple high-value targets simultaneously.
6. Rational Filtering
To refine the shortlist, apply pragmatic filters: retain only FDA-approved small molecules with drug-like properties, ensure scaffold diversity, and avoid redundancy in mechanism of action or target profile.
7. From Prediction to Validation
Predicted candidates move to experimental testing, starting with cell-based assays tailored to the disease model—such as infection inhibition assays, pseudovirus entry tests, or phenotypic screens. Hits are further confirmed with orthogonal validation techniques, from qRT-PCR to targeted inhibition assays.
By uniting network pharmacology, graph machine learning, and experimental feedback, this interactome-driven strategy offers a scalable framework for uncovering repurposing opportunities across a broad range of diseases. Its strength lies in connecting molecular context with computational inference—transforming existing drugs into tomorrow’s targeted therapies.
