DrugPath -> Biomedical Drug Reasoning Agent
Agent Name
DrugPath
What it does
DrugPath is an AI agent that navigates a biomedical knowledge graph to answer
questions about drugs, their mechanisms of action, interactions, and disease
connections. Unlike a drug database that returns flat yes/no answers, DrugPath
explains the biological mechanism behind every answer by traversing
multi-hop paths through the graph.
Three key capabilities:
1. Drug interaction checking with mechanism explanation
Instead of "warfarin + aspirin = dangerous," DrugPath explains: "These drugs
interact because warfarin inhibits VKORC1 (Vitamin K pathway) while aspirin
irreversibly blocks COX-1 - both affect different steps of the clotting cascade,
creating synergistic bleeding risk." The answer traces the path: Drug → shared
Gene targets → Pathway → clinical outcome.
2. Drug repurposing discovery
"What existing approved drugs could work for Alzheimer's?" - This requires a
3-hop traversal: find drugs → whose molecular targets (Genes) → are associated
with Alzheimer's Disease → but the drug isn't currently indicated for it. No SQL
join can express this cleanly. In the graph it's one Cypher query.
3. Comprehensive drug profiles
Full pharmacological context for any drug: indications, molecular targets,
biological pathways, side effects - all pulled from graph relationships rather
than flat attributes.
Dataset and Why a Graph Fits
Dataset: Hetionet v1.0 (see my github for link, I have limitations here) (CC0 license)
- 47,031 nodes of 11 types
- 2,250,197 relationships of 24 types in the source; we load all nodes and the
~293k relationships across the ten metaedges the agent traverses (AuraDB Free
caps at 400k relationships) - Integrates 29 public biomedical databases: DrugBank, OMIM, DisGeNET, Reactome,
Gene Ontology, SIDER, and others
Why a graph is the only right structure for this data:
A relational database can tell you: "Metformin treats Type 2 Diabetes."
A knowledge graph can answer: "Why might Metformin work against cancer?"
(Metformin:Compound)
-[:BINDS_GENE]-> (SLC22A1:Gene)
-[:ASSOCIATES_WITH]-> (prostate cancer:Disease)
This is a real path returned by the live graph. It reveals a biologically
plausible repurposing hypothesis that is not in any flat table - it emerges from
the structure of the graph itself.
Node types:
| Node | Count | Role |
|---|---|---|
| Compound | 1,552 | Drugs and chemical compounds |
| Gene | 20,945 | Molecular targets and enzymes |
| Disease | 137 | Therapeutic indications |
| Pathway | 1,822 | Biological pathways (mTOR, CYP3A4...) |
| SideEffect | 5,734 | Adverse effects |
| PharmacologicClass | 345 | Drug classes |
| Anatomy | 402 | Where diseases manifest |
Key relationships: TREATS, PALLIATES, BINDS_GENE, DOWNREGULATES_GENE,
UPREGULATES_GENE, CAUSES_SIDE_EFFECT, PARTICIPATES_IN, ASSOCIATES_WITH,
LOCALIZES_TO, INCLUDES
Agent Tools
| Tool | Type | What it enables |
|---|---|---|
drug_interaction_checker |
Cypher Template | Given two drug names, traverses shared gene targets and pathways to explain the interaction mechanism - not just flag it |
drug_repurposing_explorer |
Cypher Template | Given a disease, finds drugs approved for OTHER conditions whose molecular targets associate with this disease - pure graph reasoning |
drug_profile_lookup |
Cypher Template | Full pharmacological profile: targets, indications, side effects, pathways, drug class |
find_similar_drugs |
Similarity Search | Vector search on drug name + mechanism embeddings to find pharmacologically similar compounds |
Example Interactions
User: "Can I take warfarin and aspirin together?"
DrugPath: "These two drugs share several molecular targets in the graph -
CYP2C9, CYP2C8, CYP2C19 (the enzymes that metabolise warfarin) and ALB
(plasma albumin) - and overlap in the arachidonic-acid metabolism pathway.
That means aspirin can compete for warfarin's metabolism and displace it from
albumin, raising free warfarin levels, while also acting on the clotting cascade
itself - a combination associated with increased bleeding risk. Path traversed:
Warfarin → shared Genes (CYP2C9, ALB) ← Aspirin. This information is for
educational purposes only - always consult a healthcare professional."
User: "What approved drugs might work for Alzheimer's disease?"
DrugPath: "Here are compounds whose molecular targets associate with
Alzheimer's disease in the graph but which are not currently indicated for it:
- Dextromethorphan - via the nicotinic receptors CHRNA4 / CHRNA7 / CHRNB2
- Nicotine - via the cholinergic targets CHAT, CHRNA4, CHRNA7
- Amitriptyline - via the serotonergic targets HTR2A, HTR6, SLC6A4
The cholinergic and serotonergic targets are biologically on-point for
Alzheimer's. Path: Drug → [BINDS_GENE] → Gene → [ASSOCIATES_WITH] → Alzheimer's
disease. These are research hypotheses, not proven treatments. Educational
purposes only."
Screenshots
Graph schema (Neo4j Browser → CALL db.schema.visualization()):
All 47,031 nodes by type (Neo4j Browser results overview):
Agent in action - full drug profile (Aura Console):
See github readme
Agent in action - similarity search (Aura Console):
See github readme
Aura dashboards (graph statistics and the compound · disease · gene network):
Also see github readme
Live Agent
Landing page: DrugPath — the drug agent that explains why
Published agent endpoint: [will be added after receiving credits]
What Makes DrugPath Different
1. The graph drives every answer, not just stores data.
Every response cites the traversal path: Compound → Gene → Pathway → Disease.
Users see why the graph gives this answer, not just what it returns.
2. Drug repurposing is impossible without the graph.
Finding drugs whose targets associate with a disease they're not indicated for
requires crossing node types in a single query. This is where
graph databases outperform anything else.
3. 29 databases, one coherent graph.
Hetionet integrates DrugBank, OMIM, DisGeNET, Reactome, Gene Ontology, and SIDER
into a single traversable structure. The agent reasons across all of them
simultaneously.
4. Real educational value.
The agent may be genuinely useful for medical students, researchers, and anyone
trying to understand pharmacology beyond "take this pill for that disease."
Technical Stack
- Graph database: Neo4j Aura (AuraDB)
- Dataset: Hetionet v1.0 (CC0)
- ETL: Python (neo4j driver, pandas)
- Embeddings: OpenAI text-embedding-3-small, matched to Aura Agent's managed similarity search
- Agent: Neo4j Aura Agent with 3 Cypher Templates + 1 Similarity Search tool
- Data loading: ~47k nodes, ~293k relationships (filtered to the ten
metaedges the agent traverses, within free limits)
DrugPath is an educational and research tool. It does not provide
medical advice. You should always consult a qualified healthcare professional for medical
decisions.