AI-powered knowledge graph mining — discovering new treatments for existing drugs using multi-omics & network pharmacology
Drug repurposing — finding new uses for existing approved drugs — is one of the most promising strategies in pharmaceutical research. It reduces drug development timelines from 12–15 years to 3–7 years and cuts costs from ~$2.5 billion to $300–600 million per approval. The key insight: drugs affect networks of biological interactions, not individual targets, and those networks are largely shared across diseases.
A biomedical knowledge graph (KG) is a heterogeneous directed graph G = (V, E, R) where: • V = nodes representing biological entities (drugs, genes, proteins, diseases, pathways, anatomy, side effects, symptoms) • E = edges representing relationships (treats, targets, upregulates, associates, participates_in, etc.) • R = edge types (encoded as relation-specific weight or embedding)
Hetionet v1.0 data sources (2020): • Compounds: DrugBank 5.0 (2,229 small molecules + biologics); NPC; ChEMBL 25 • Genes/Proteins: Entrez Gene + UniProt (20,945 human genes; 17,015 eligible proteins) • Diseases: DOID (Disease Ontology) — 137 curated complex diseases • Pathways: Reactome 5.3, WikiPathways, KEGG (1,822 pathways) • Gene interactions: STRING/BioGRID/DIP (physical protein-protein interactions) • Drug-target: BindingDB IC50 <1μM filter; ChEMBL curated bioactivities • Disease-gene: DisGeNET (curated genetic associations); OMIM; ClinVar • Side effects: SIDER (drug label mining, 5,868 compounds × 5,734 side effects) • Literature: PubMed 28M abstracts → co-mention mining (ScispaCy NLP)
Edge construction quality tiers: • Tier 1 (high confidence): FDA label binding data, crystal structures, clinical trial records • Tier 2 (medium): curated databases (ChEMBL IC50 <1μM, UniProt functional annotation) • Tier 3 (computationally predicted): text mining with confidence >0.8, homology transfer, ML-predicted interactions • Quality filter: minimum 2 independent evidence sources for inclusion in final KG
Graph storage and query infrastructure: • Neo4j: native graph database; Cypher query language • Example query: MATCH (d:Drug)-[:TARGETS]->(p:Protein)-[:PARTICIPATES_IN]->(pw:Pathway) RETURN d, pw LIMIT 100 • RDF/SPARQL: OWL ontologies enable semantic reasoning (transitive subclass inference) • Scale: 2.25M edges → Neo4j storage: 4.8 GB; full-graph query time: <2 seconds on SSD
A drug's therapeutic potential for a disease is encoded in the topological relationship between the drug's targets and the disease's associated genes in the protein-protein interaction network. This "network medicine" insight, formalized by Barabási and Menche (2015), provides a principled way to predict which drugs are likely to modulate a disease — independent of mechanistic understanding.
Random Walk with Restart (RWR) algorithm:
Setup: • Adjacency matrix A of the protein-protein interaction network (19,000 nodes, 230,000 edges) • Normalize: W = D⁻¹A where D is degree matrix • Drug seed set S_d = {nodes} for drug d (known protein targets of drug) • Restart probability r = 0.2 (typical value)
Iteration: • p_{t+1} = (1-r) × W × p_t + r × p_0 • p_0 = initial probability vector (1/|S_d| at each seed node, 0 elsewhere) • Converge when ||p_{t+1} - p_t|| < 10⁻⁶ • Result: p_∞ = proximity score of every node to the drug-target subnetwork
Disease scoring: • Score(drug d, disease D) = mean(p_∞ for disease gene set G_D) • Z-score normalization: compare to random seed sets of same size • P-value: fraction of 10,000 random seed sets scoring ≥ actual score • Correction: Benjamini-Hochberg FDR < 0.05 for reporting
Network separation score (Menche 2015 Science): • s_AB = <d_AB> - (<d_AA> + <d_BB>)/2 where <d_AB> = mean shortest path between drug A and disease B modules • s_AB < 0: overlap or closeness → drug likely active in disease • s_AB > 0: separation → drug unlikely to affect disease • Validated on 402 drug-disease pairs with known clinical outcomes • Accuracy: s_AB < 0 predicts known indication in 71% of cases vs. 22% expected by chance
Key repurposing examples from network proximity: • Sildenafil (erectile dysfunction, PDE5 inhibitor): network proximity to pulmonary arterial hypertension → approved as Revatio in 2005 • Metformin (T2DM, AMPK activator): network proximity to cancer metabolism network → 120+ ongoing cancer trials • Baricitinib (rheumatoid arthritis, JAK1/2 inhibitor): RWR proximity to COVID-19 cytokine storm nodes → FDA EUA 2020 → Phase III validated
Traditional methods compute similarity between drugs and diseases based on pre-specified similarity metrics. Graph Neural Networks (GNNs) learn distributed representations of all entities simultaneously by propagating information across the graph structure. This allows the model to discover complex, multi-hop relationships that no human-curated rule could capture — and to generalize to novel drug-disease pairs never seen during training.
Knowledge Graph Embedding (KGE) models learn entity and relation embeddings such that true triples have low energy:
TransE (Bordes 2013 — foundational model): • Embedding: h + r ≈ t for true triple (head, relation, tail) • Score: -||h + r - t||₂ (lower = more likely) • Limitations: cannot model symmetric (A interacts B ↔ B interacts A) or 1-to-N relations • Useful for: simple drug-target prediction, disease-gene associations
RotatE (Sun 2019): • Models each relation as rotation in complex embedding space: h ⊙ r = t • Handles: symmetry, antisymmetry, inversion, composition — full set of relation patterns • Performance: DRKG (Drug Repurposing Knowledge Graph): AUC 0.918 on validation set
R-GCN (Schlichtkrull 2018 — Relational Graph Convolutional Network): • For node v: h_v^(l+1) = σ(W_0^l × h_v^l + Σ_r Σ_{u∈N_r(v)} W_r^l × h_u^l / |N_r(v)|) • W_r^l: learnable weight matrix per relation type r • Enables: node classification (disease subtype), link prediction (drug-disease), graph classification • Applied to: COVID-19 repurposing (DRKG 5.8M triples); Alzheimer's (3 papers each using R-GCN)
DRKG (Zeng 2020, IBM/DGL): • Drug Repurposing Knowledge Graph built during COVID-19 pandemic • 97,238 entities (drugs, genes, diseases, biological processes, anatomy, side effects) • 5,874,261 relations from GNBR, Hetionet, STRING, DrugBank, IntAct, DGIdb • RotatE trained on DRKG → ranked repurposing candidates for SARS-CoV-2 • Top candidates: baricitinib (#3), dexamethasone (#12), remdesivir (#1) • All 3 received FDA EUA/approval within 8 months of graph construction
Tip: combining GNN embedding + network proximity in an ensemble improves AUC from 0.89 to 0.94 on benchmark datasets — different methods capture complementary evidence
The Connectivity Map (CMap), conceived by Broad Institute in 2006 and massively scaled to L1000 in 2017, is based on a powerful idea: if a disease upregulates gene X, then a drug that downregulates gene X might treat the disease. By building a library of gene expression profiles for thousands of drugs and then computing the "connectivity" between a disease signature and drug signatures, we can identify drugs that are likely to reverse the disease state at the transcriptomic level.
L1000 high-dimensional transcriptomics at scale:
Technology: • Luminex bead-based multiplexed assay: 978 color-coded bead pools, each capturing one landmark gene mRNA via hybridization • Cells treated with drug/siRNA/CRISPR/overexpression → measure 978 landmark genes • Missing 10,000 genes inferred by linear regression trained on LINCS 2 full RNA-seq reference (R² = 0.82 median) • Throughput: 384-well plates; $8/sample at full scale; 1.5M profiles as of 2022
Connectivity score calculation (GSEA-based): 1. Input: disease signature = ranked list of differentially expressed genes (e.g., Alzheimer's patient brain RNA-seq vs. control) 2. Query UP-genes (upregulated in disease), DOWN-genes (downregulated) separately 3. For each L1000 compound profile: a. Rank all 978 landmark genes by change magnitude in drug treatment b. Compute enrichment score (KS-like statistic) for disease UP-genes in drug-DOWN ranked list c. Compute enrichment score for disease DOWN-genes in drug-UP ranked list d. Combined score: τ = (ES_UP_in_DOWN + ES_DOWN_in_UP) / 2 4. τ ∈ [-100, +100]: τ < -85 = strong reversal (repurposing candidate); τ > +85 = strong concordance (potential disease inducer) 5. Normalize by cell line: score within cell line to remove context-specific expression effects
Historical validation cases: • Imatinib GI stromal tumor (GIST): imatinib originally kinase inhibitor for CML; CMAP connectivity score high between imatinib profile and GIST signature → predicted 2001; confirmed 2002 FDA approval • Topiramate for binge eating disorder: connectivity score with hypothalamic appetite regulation signature → randomized trial confirmed → FDA approval 2023 • Multiple sclerosis: CMAP identified cladribine as reversal drug for MS gene signature 3 years before clinical validation • COVID-19: imatinib scored -92 against COVID-19 severe disease signature → clinical trial showed reduced ICU time
Current limitations: • Cell lines ≠ primary patient cells; expression reflects context • 978 genes only partially capture pathway activity • Off-target effects (toxicity, metabolism) not encoded in signatures • Solution: patient-derived organoids express same or better connectivity scores → patient-matched drug screening
No single computational method is sufficient to identify reliable repurposing candidates. The most successful programs combine evidence from network proximity, GNN embeddings, transcriptomic connectivity, literature mining, and real-world clinical signals (electronic health records, pharmacovigilance data) into an integrated scoring framework — then validate the top candidates experimentally before advancing to clinical trials.
Electronic Health Record (EHR) mining for repurposing signals:
PheWAS (Phenome-Wide Association Study): • Query: patients prescribed drug X — what is their incidence of disease Y vs. matched controls? • Dataset: TriNetX (100M+ patients), CPRD (UK, 16M patients), FDA FAERS • Statistical model: logistic regression; OR < 0.5, p < 0.00001 after Bonferroni correction = signal • Example: patients on GLP-1 agonists (liraglutide, semaglutide) show 50% lower incidence of dementia → motivates NeuroBiological GLP-1R Alzheimer's trial • Metformin PheWAS: significant inverse association with 12 cancer types, heart failure, and macular degeneration
Real-World Evidence (RWE) signal types: • Protective signal: patients on drug X have lower incidence of disease Y • Confounder risks: indication bias, healthy user bias, detection bias • Causal inference controls: propensity score matching, instrumental variable analysis (Mendelian randomization) • Gold standard: Mendelian randomization uses genetic variants as instrument to simulate "natural randomized trial"
Patient-derived organoid (PDO) validation: • Tumor organoids from patient biopsies grown in matrigel or AIR (air-liquid interface) • Drug sensitivity profile: IC50, EC50 measured across 200–500 compound library • 3D structure enables pharmacokinetic access testing (drug penetration into organoid center) • Clinical correlation: organoid drug sensitivity predicts patient response in >65% of Phase II trials (Vlachogiannis 2018 Science) • Repurposing workflow: 1000 computational candidates → 50 organoid-screened → 5 clinical candidates
Composite evidence scoring framework: • E1: RWR network proximity z-score • E2: GNN/RotatE embedding similarity score • E3: CMAP transcriptomics connectivity τ score • E4: Literature co-occurrence (PubMed mining, BERT-based NER) • E5: EHR protective odds ratio (if available) • E6: Structural similarity to known indication drugs (Tanimoto coefficient) • Final score: F = weighted sum (learned by logistic regression on training set with known outcomes) • Public tool: OpenTargets, DGIdb, CBioPortal integrate many of these evidence streams
The COVID-19 pandemic became the largest-ever stress test of computational drug repurposing. Within 3 months, 50+ artificial intelligence groups published repurposing predictions from knowledge graphs, GNNs, docking screens, and CMAP analysis. From this flood of predictions, baricitinib (JAK1/2), dexamethasone (corticosteroid), and tocilizumab (IL-6 receptor antibody) were validated in large randomized trials. All three had been flagged by AI-based knowledge graph models by March 2020 — before clinical trial results were available in June–October 2020.