DrugBank

DrugBank organizes data as a heterogeneous biomedical knowledge graph where drug entities are central nodes connected to targets, enzymes, transporters, and carriers. This graph structure enables link prediction algorithms to identify novel drug-disease associations by traversing relationships between a drug's known protein targets and disease-associated genes. The graph contains over 15,000 drug entries with structured data on chemical structures, pharmacokinetics, and pharmacodynamics, making it the gold standard for training knowledge graph embedding models like TransE and RotatE for drug repurposing.

Each drug entry contains structured indication fields that explicitly map drugs to approved therapeutic uses, along with curated off-label applications. This structured annotation serves as the ground truth for training multi-task learning models that simultaneously predict both approved and novel indications. The indication data is organized hierarchically, linking drugs to specific disease concepts through controlled vocabularies, which enables zero-shot prediction of repurposing candidates for rare diseases that lack dedicated treatment options.

DrugBank provides a comprehensive binary drug-target interaction matrix where rows represent drugs and columns represent protein targets, with entries indicating known binding relationships. This matrix is the primary input for matrix factorization and inductive matrix completion algorithms that decompose the sparse interaction landscape into latent factor representations. By integrating additional side information such as molecular fingerprints and protein sequence features, these models can predict binding affinities for completely uncharacterized drug-target pairs, directly informing polypharmacology studies.

DrugBank integrates pharmacogenomic annotations that document how genetic variations affect drug response, metabolism, and toxicity. This includes structured data on cytochrome P450 enzyme interactions, transporter polymorphisms, and target gene variants. For computational drug repurposing, these annotations enable models to stratify patient populations and predict adverse drug reactions based on genetic profiles. The pharmacogenomic layer transforms DrugBank from a simple drug database into a precision medicine resource that supports causal inference studies linking genetic variants to drug outcomes.

DrugBank employs a multi-level chemical taxonomy system that classifies drugs by chemical structure, mechanism of action, and therapeutic category. This hierarchical classification enables drug similarity network construction where compounds sharing the same taxonomic branches are likely to exhibit similar biological activities. The taxonomy supports contrastive learning approaches by providing natural positive pairs—drugs within the same class—and negative pairs across different classes, allowing models to learn robust molecular representations without explicit binding data.

DrugBank contains detailed drug metabolism pathways documenting how prodrugs are converted to active metabolites and how drugs are eliminated. Each drug entry maps to specific metabolic enzymes and transporter proteins, creating a comprehensive view of a drug's journey through the body. This metabolic network is critical for side effect prediction models that identify off-target effects caused by reactive metabolites or drug-drug interactions at shared metabolic pathways. The pathway data also supports network propagation algorithms that diffuse drug signals through biological networks.

DrugBank's structured data on drug-target interactions, mechanisms of action, and pharmacokinetics serves as the backbone for constructing biomedical knowledge graphs. These graphs represent drugs, proteins, diseases, and pathways as nodes, with edges defining their relationships. Graph neural networks then perform link prediction to infer novel drug-disease associations for repurposing.

• Provides high-quality, curated edges for heterogeneous graph initialization.
• Enables polypharmacology modeling by mapping a single drug to multiple protein targets.
• Reduces reliance on noisy, unstructured text mining for relationship extraction.

The database provides canonical SMILES strings, InChI keys, and detailed chemical taxonomy for every entry. These structured chemical identifiers are the direct input for calculating molecular fingerprints (e.g., ECFP4, MACCS) and physicochemical descriptors used in Quantitative Structure-Activity Relationship (QSAR) models.

• Enables the generation of fixed-length vector representations for machine learning.
• Facilitates drug similarity network analysis by providing standardized chemical structures.
• Supports contrastive learning by defining positive and negative molecular pairs based on structural similarity.

DrugBank's comprehensive mapping of drugs to their primary and off-target proteins is critical for target deconvolution—the process of identifying the molecular target responsible for an observed phenotypic effect. Computational methods use this data to train models that predict binding profiles for novel compounds.

• Lists known carriers, transporters, and enzymes modulated by each drug.
• Provides the ground truth for training drug-target interaction prediction models.
• Essential for distinguishing therapeutic effects from side effect prediction signals.

DrugBank entries are often cross-referenced with resources like the Connectivity Map (CMap). By linking a drug's known pharmacological profile to its induced gene expression signature, researchers can perform transcriptomic signature matching. This identifies drugs that reverse a disease's gene expression pattern, a core technique in computational drug repurposing.

• Bridges the gap between chemical structure and systems biology response.
• Provides the metadata required to interpret high-throughput screening results.
• Enables zero-shot prediction of drug efficacy for diseases with known transcriptomic profiles.

The database's structured data on adverse drug reactions and drug-food interactions provides a gold standard for training side effect prediction models. Machine learning algorithms use this data to correlate chemical substructures with specific toxicities, enabling proactive safety profiling for repurposed candidates.

• Contains structured data on adverse reactions and toxicity thresholds.
• Supports the development of multi-task learning models that simultaneously predict efficacy and safety.
• Used to validate predictions from Electronic Health Record Mining against known pharmacological mechanisms.

DrugBank's detailed mechanism of action data is essential for drug combination prediction. Algorithms analyze the pathways and targets of multiple drugs to predict synergistic effects, where the combined therapeutic impact exceeds the sum of individual effects. This is quantified using a synergy score.

• Maps drugs to their specific mechanism of action (MoA) for pathway-level analysis.
• Enables computational screening for synergistic pairs that target parallel or compensatory pathways.
• Provides the foundational data for models that avoid antagonistic combinations.

The systematic identification of new therapeutic indications for existing, clinically approved drugs outside the scope of their original medical use. DrugBank serves as the gold-standard reference for this process by providing structured data on approved compounds.

• Mechanism: Matches known drug-target profiles to novel disease pathways
• Advantage: Significantly reduces development timelines and failure risks compared to de novo discovery
• Example: Identifying that a drug approved for hypertension also inhibits a kinase implicated in a rare cancer

The physical binding event between a chemical compound and a specific biomolecule, typically a protein, which modulates a biological process. DrugBank meticulously curates these binary interactions to power predictive models.

• Data Types: Includes binding affinities (Ki, Kd, IC50), mechanism of action, and functional effects
• Computational Use: Forms the training labels for graph neural networks predicting novel off-targets
• Polypharmacology: A single drug often has multiple targets; DrugBank maps these complex profiles

A machine learning technique that projects the entities and relations of a biomedical knowledge graph into a low-dimensional vector space. DrugBank's structured data is often converted into a heterogeneous graph to predict missing links.

• Nodes: Drugs, proteins, diseases, pathways, and side effects
• Edges: 'binds_to', 'treats', 'causes', 'metabolizes'
• Outcome: Generates high-probability drug-disease association predictions for experimental validation

The design or functional capacity of a single drug molecule to interact with multiple distinct biological targets simultaneously. DrugBank's comprehensive target profiles enable the intentional design of multi-target drugs.

• Therapeutic Polypharmacology: A single drug hitting multiple nodes in a disease network for synergistic efficacy
• Adverse Polypharmacology: Binding to off-targets causing toxicity; predicted using DrugBank's side-effect data
• Kinase Inhibitors: A classic example where promiscuity is often therapeutically beneficial

The computational forecasting of adverse drug reactions by modeling the interaction between a drug's chemical structure and off-target biological pathways. DrugBank provides the ground-truth labels for training these safety models.

• Input Features: Chemical substructures, protein target profiles, and transcriptomic signatures
• Modeling Approach: Often framed as a multi-label classification problem using deep neural networks
• Clinical Impact: Early identification of cardiotoxicity or hepatotoxicity risks before preclinical testing

A computational approach that compares a disease's gene expression pattern against a reference database of drug-induced gene expression changes. DrugBank links drugs to their known perturbation signatures in resources like the Connectivity Map.

• Logic: If a drug's signature inversely correlates with a disease signature, it is a repurposing candidate
• Data Integration: Combines DrugBank's structural data with high-throughput transcriptomic readouts
• Application: Identifying candidates that reverse the molecular pathology of neurodegenerative diseases