Drug Reasoning API
A drug reasoning API that explains how drugs, targets, pathways, and symptoms connect — using graph-based causal inference and real biomedical sources.
Reasoning across 1.6 million+ biomedical relationships from 7 databases
Knowledge Graph Scale
The reasoning layer medical AI has been missing
Most medical AI retrieves information or predicts patterns. This system traces mechanistic relationships across drugs, targets, pathways, and effects.
Graph-Based Causal Chains
Builds a knowledge graph from ChEMBL, Reactome, CTD, and FAERS. Finds causal paths via graph traversal — not LLM generation.
Real Data, Full Provenance
Every claim traces back to a specific database, ID, and confidence score. No hallucinated references.
LLM as Explanation Layer
The LLM receives graph-derived data and generates a human-readable summary. It never invents relationships or sources.
How We Compare
Not Another Medical Chatbot
| Capability | Drug Reasoning API | Symptom Checkers | LLM Chatbots | Lit. Search |
|---|---|---|---|---|
| Mechanistic reasoning | ||||
| Causal chain output | ||||
| Provenance & citations | Partial | |||
| Structured API output | ||||
| Confidence scoring | ||||
| Graph-based reasoning | ||||
| Real-time source attribution | Partial | |||
| Low hallucination risk |
Live Query
Try the Reasoning Engine
Ask about drug mechanisms, side effects, or pathways. Every result is graph-derived with full source attribution.
Try these questions
Sample Result
What the API Returns
A complete reasoning result — normalized entities, causal chain, explanation, confidence, and provenance.
Interactive Causal Graph
Causal Chain
Plain-Language Explanation
Confidence: 88%Lisinopril inhibits angiotensin-converting enzyme (ACE), which normally degrades bradykinin. When ACE is inhibited, bradykinin accumulates in the pulmonary tissues, stimulating sensory nerve fibers in the airways via prostaglandin and substance P pathways. This sensitization of the cough reflex produces the characteristic dry, persistent cough reported in 5–35% of patients taking ACE inhibitors.
Confidence is calculated from source reliability, evidence density, and path strength across the causal chain.
LLM explains graph-derived data. It does not invent claims.
Data Sources & Provenance
Normalized Entities
For Developers
REST API
Structured JSON responses with normalized entities, causal chains, confidence scores, and full provenance.
Endpoints
/v1/drugs/{drug}/mechanismFull mechanistic analysis: targets, pathways, causal chains, explanation.
/v1/drugs/{drug}/side-effects/{symptom}Causal reasoning chain from drug to a specific side effect.
/v1/entities/search?q={query}Search normalized biomedical entities across drugs, proteins, pathways.
/v1/healthService health check with upstream data source status.
Quick Start
curl -s "https://bmdrm.site/v1/drugs/lisinopril/mechanism" | jq .Sample Response
{
"drug": {
"name": "Lisinopril",
"id": "CID:5362119",
"formula": "C21H31N3O5",
"source": "PubChem"
},
"targets": [
{
"name": "Angiotensin-converting enzyme",
"gene": "ACE",
"uniprot": "P12821",
"action": "INHIBITOR",
"source": "ChEMBL",
"confidence": 0.95
}
],
"causalChains": [
{
"chain": [
{ "entity": "Lisinopril", "type": "drug" },
{ "relation": "INHIBITS", "source": "ChEMBL", "confidence": 0.95 },
{ "entity": "ACE", "type": "protein" },
{ "relation": "PARTICIPATES_IN", "source": "Reactome", "confidence": 0.88 },
{ "entity": "Bradykinin signaling", "type": "pathway" },
{ "relation": "ASSOCIATED_WITH", "source": "CTD", "confidence": 0.82 },
{ "entity": "Cough", "type": "side_effect" }
],
"overallConfidence": 0.88
}
],
"explanation": "Lisinopril inhibits ACE, causing bradykinin accumulation...",
"provenance": [
{ "source": "ChEMBL", "id": "CHEMBL1808", "url": "https://www.ebi.ac.uk/chembl/target/CHEMBL1808" },
{ "source": "CTD", "pmids": ["15947090", "18784654"] },
{ "source": "OpenFDA FAERS", "reportCount": 34219 }
]
}Evidence Layer
Data Sources & Provenance
Every claim is backed by a specific database, identifier, and confidence score. No data is fabricated or LLM-generated.
PubChem
Drug identity & properties
111M+ compounds
Updated weekly
ChEMBL
Target binding data
2.4M+ bioactivity records
Quarterly releases
UniProt
Protein function & structure
20,400+ human proteins
Monthly releases
Reactome
Biological pathways
2,600+ human pathways
Quarterly releases
CTD
Chemical-gene-disease relationships
1.6M+ curated interactions
Monthly updates
OpenFDA FAERS
Adverse event reports
28M+ reports
Quarterly updates
Open Targets
Target-disease associations
14,000+ target-disease pairs
Biannual releases
DrugMechDB
GOLD STANDARDGold-standard validation benchmark
4,500+ curated mechanism paths
Validation benchmark (not queried live)
Provenance Guarantee
Every relationship includes source database, identifier, evidence note, and confidence score. The LLM only explains — it never invents data.
Who It's For
Built for Builders
A reasoning infrastructure layer for developers, research teams, and health AI systems.
Health AI Platforms
Add mechanistic drug reasoning to patient-facing or clinical AI systems without building a knowledge graph from scratch.
"A clinical decision support tool queries the API to explain why a prescribed drug might cause a reported symptom."
Biomedical Research
Explore drug-target-pathway-effect relationships with full provenance and confidence scores for hypothesis generation.
"A research team uses batch queries to map all known mechanistic pathways for a set of candidate compounds."
Developer Integration
Consume structured JSON via REST endpoints. Get normalized entities, causal chains, and confidence scores programmatically.
"A health app integrates the mechanism endpoint to show patients why their medication has specific side effects."
Drug Safety & Pharmacovigilance
Trace mechanistic pathways from drug to adverse event with supporting evidence from CTD, FAERS, and ChEMBL.
"A safety team traces the mechanistic pathway from a drug to a newly reported adverse event using graph-derived evidence."
Architecture
How It Works
Normalize
Drug name resolved via PubChem. Brand names mapped to generic. Symptoms normalized to MedDRA terms.
Build Graph
Fetch targets (ChEMBL), pathways (Reactome), gene interactions (CTD), adverse events (FAERS). Build in-memory knowledge graph.
Find Chains
Traverse graph to find causal paths from drug → target → pathway → effect. Score each chain by combined confidence.
Explain
LLM receives graph data + provenance. Generates plain-language explanation. Never invents relationships.
The knowledge graph is constructed per request from live upstream APIs. Typical response time is 15–45 seconds for uncached queries. Results are cached in PostgreSQL so subsequent queries for the same drug return in under 1 second.
Pricing
Simple, Scalable Pricing
Start free. Scale when ready.
Pro
- 2,000 queries/month
- All endpoints
- Priority support
- Higher rate limits (30/min)
- Batch query access
Enterprise
- Unlimited queries
- Custom integrations
- Dedicated support & SLA
- Custom data sources
- On-premise option
Accuracy
Evaluation Results
Methodology
We validated the engine against 42 well-documented drugs whose accepted primary targets are curated in DrugMechDB and cross-checked against pharmacology references. For each drug, the engine reconstructs targets from the live pipeline (ChEMBL mechanism of action → UniProt) and is matched to the gold standard by canonical UniProt accession — not by name strings, which eliminates false matches from gene-symbol collisions (e.g. ACE / P12821 vs ACE2 / Q9BYF1). A drug is scored correct if the engine recovers at least one accepted target accession.
The engine reconstructed the correct canonical target for 41 of 42 drugs (97.6%). The single mismatch is metformin: ChEMBL curates its direct molecular target as mitochondrial complex I, while the clinically-cited effector is AMPK (PRKAB1) downstream. The commonality across residual mismatches is indirect/diffuse mechanisms where the curated direct binding target differs from the downstream clinical effector — a known mechanism ambiguity, not an identifier-resolution error.
Reproducible harness — every number above is computed live from the data sources, not hand-entered. See the full validation table.
Early Access
Request API Access
Join the waitlist to get your API key when access opens.
Not another medical chatbot.
A mechanism-aware drug reasoning layer.
Graph reasoning over real databases. Structured outputs with provenance. Built for developers and biomedical AI systems.