Drug Similarity Network

Each node in the network corresponds to a distinct drug entity, typically identified by its DrugBank ID or PubChem CID. Nodes are annotated with rich metadata including:

• Chemical structure (SMILES, InChI)
• Molecular fingerprints (ECFP4, MACCS keys)
• Known protein targets and their binding affinities
• ATC classification codes for therapeutic categorization
• Side effect profiles from pharmacovigilance databases like SIDER

This multi-view annotation allows the same drug to participate in multiple similarity networks simultaneously, each capturing a different biological dimension.

Edges between drug nodes are weighted by quantitative similarity metrics. Common edge-weighting strategies include:

• Tanimoto coefficient for 2D chemical fingerprint overlap (range: 0 to 1)
• Semantic similarity between Gene Ontology terms of shared protein targets
• Cosine similarity between side effect frequency vectors extracted from FAERS
• Transcriptomic correlation between gene expression signatures from the Connectivity Map

A high edge weight indicates a strong likelihood of shared mechanism of action or therapeutic interchangeability. Thresholding is often applied to sparsify the graph and reduce noise.

Advanced drug similarity networks are heterogeneous graphs that incorporate multiple node types beyond drugs:

• Protein targets linked via drug-target interaction edges
• Disease nodes connected through drug-disease association edges
• Gene nodes linked via drug-induced differential expression edges
• Side effect nodes connected through drug-adverse event edges

This heterogeneous structure enables meta-path-based similarity, where two drugs are considered similar if they share paths through intermediate biological entities, even without direct chemical similarity. Graph neural networks like HAN and HGT exploit this topology for link prediction.

Modularity-based clustering algorithms such as the Louvain method or Leiden algorithm partition the drug similarity network into densely connected communities. These clusters often correspond to:

• Drugs sharing a therapeutic class (e.g., beta-blockers)
• Compounds targeting the same protein family (e.g., kinase inhibitors)
• Molecules with a common chemical scaffold (e.g., benzodiazepines)

Community detection enables drug repositioning by identifying approved drugs that cluster with investigational compounds for a novel indication, suggesting shared efficacy. Overlapping community detection methods like Clique Percolation capture drugs with polypharmacological profiles that belong to multiple therapeutic clusters.

Random walk with restart (RWR) is the foundational algorithm for propagating information across drug similarity networks. Starting from a set of seed nodes (known drugs for a disease), RWR diffuses probability mass through weighted edges, producing a stationary distribution that ranks all nodes by proximity to the seeds.

Variants include:

• Heat kernel diffusion for capturing local neighborhood effects
• PageRank-inspired algorithms that incorporate node degree priors
• Multi-layer propagation across stacked similarity networks (chemical, biological, clinical)

These algorithms power guilt-by-association predictions, identifying drugs that are topologically proximal to known effective treatments for a given indication.

Modern drug similarity networks are processed using Graph Neural Networks (GNNs) that learn low-dimensional vector embeddings for each drug node. Architectures include:

• Graph Convolutional Networks (GCNs) that aggregate features from local neighborhoods
• Graph Attention Networks (GATs) that learn edge-specific attention weights
• GraphSAGE for inductive learning on unseen drugs via neighbor sampling

These embeddings capture both topological context and node attributes, enabling downstream tasks such as drug-drug interaction prediction, side effect forecasting, and polypharmacology modeling. Contrastive learning objectives further refine embeddings by maximizing agreement between augmented views of the same drug's neighborhood.

A binary or integer vector encoding the presence or absence of specific chemical substructures within a molecule. These fixed-length representations serve as the primary input for calculating Tanimoto similarity to construct chemical similarity edges in the network.

• MACCS Keys: 166-bit structural key-based fingerprint
• Morgan Fingerprints: Circular fingerprints capturing atom neighborhoods
• ECFP4: Industry-standard extended-connectivity fingerprint for structure-activity modeling

A metric quantifying the structural overlap between two molecular fingerprints, defined as the ratio of shared bits to total bits. It is the most common edge-weighting function in drug similarity networks.

• Formula: J(A,B) = |A ∩ B| / |A ∪ B|
• Threshold: Typically >0.7 indicates significant structural similarity
• Limitation: Does not capture bioisosterism or 3D pharmacophoric similarity

A high-dimensional vector cataloging the frequency and severity of adverse drug reactions extracted from FAERS or SIDER databases. Drugs are connected in the network if they share statistically significant side effect signatures, often revealing a common off-target mechanism.

• Cosine similarity is typically used to compare profiles
• Enables identification of therapeutic class polypharmacology
• Used to predict novel adverse events for understudied compounds

A genome-wide gene expression profile induced by drug treatment, typically derived from the Connectivity Map (CMap) or LINCS L1000 platform. Nodes are connected if their signatures are highly correlated or anti-correlated.

• Gene Set Enrichment Analysis (GSEA) quantifies pathway-level similarity
• Anti-correlated signatures suggest therapeutic reversal of disease states
• Enables mechanism-agnostic repurposing hypotheses

A technique that projects heterogeneous biomedical entities and relations into a low-dimensional continuous vector space. In a drug similarity network, embeddings capture latent properties that transcend explicit structural or phenotypic similarity.

• TransE and RotatE model relational patterns
• ComplEx handles asymmetric relations in complex space
• Enables link prediction for novel drug-disease associations

An algorithm that diffuses information across the graph's topology to prioritize nodes. A random walk with restart from known drug targets smooths association signals over the similarity network to identify disease modules.

• Random Walk with Restart (RWR) is the canonical approach
• Heat kernel diffusion provides an alternative smoothing function
• Identifies drugs that are topologically proximal to disease-associated proteins