Methodology

From AI prediction to evidence validation

---

Overall Pipeline

TxGNN Prediction → Data Collection → Evidence Integration → Report Generation

---

Step 1: TxGNN Prediction

Model Introduction

TxGNN uses a knowledge graph combined with graph neural networks for prediction:

1. Knowledge Graph Construction
   • 17,080 nodes (drugs, diseases, genes, etc.)
   • Complex inter-node relationships
2. Graph Neural Network Training
   • Learns latent associations between nodes
   • Predicts new drug-disease relationships
3. Prediction Output
   • Prediction score for each drug-disease pair
   • Higher scores indicate higher confidence

Prediction Methods

Method	Description	Predictions
Knowledge Graph (KG)	Network topology-based associations	718
Deep Learning (DL)	Neural network score prediction	31,650
Total	Merged and deduplicated	32,368

---

Step 2: Drug Mapping

EMA to DrugBank Mapping

EMA medicines are mapped to DrugBank identifiers for TxGNN compatibility:

Step	Description
1. Ingredient Extraction	Extract active substances from EMA data
2. Name Normalization	Standardize drug names (remove salts, etc.)
3. DrugBank Lookup	Match against DrugBank vocabulary
4. Manual Review	Verify ambiguous mappings

Mapping Statistics

Metric	Value
Total EMA Drugs	1,543
Mapped to DrugBank	638
Mapping Rate	41.3%

---

Step 3: Disease Mapping

Indication Extraction

Therapeutic indications from EMA are mapped to standardized disease terms:

1. MeSH Terms: EMA provides MeSH therapeutic areas
2. Text Extraction: Parse indication text for disease mentions
3. Vocabulary Matching: Match against TxGNN disease vocabulary

Mapping Statistics

Metric	Value
Unique Indications	4,570
Mapped Diseases	~2,000

---

Step 4: Evidence Levels

Classification System

Level	Definition	Clinical Significance
L1	Multiple Phase 3 RCTs / Systematic Reviews	Strong support
L2	Single RCT or multiple Phase 2 trials	Moderate support
L3	Observational studies / Large case series	Preliminary evidence
L4	Preclinical / Mechanistic studies	Reasonable mechanism
L5	AI prediction only	Research hypothesis

---

Data Sources

Source	Usage
EMA Medicines Database	EU approved drugs
Article 57 Database	Additional drug data
TxGNN Knowledge Graph	Drug-disease predictions
DrugBank	Drug identifiers
ClinicalTrials.gov	Clinical evidence
PubMed	Literature evidence

---

Limitations

• Predictions require clinical validation
• Limited to drugs with DrugBank mappings
• Based on TxGNN model trained primarily on US data
• Evidence levels require manual curation

---

Citation

@article{huang2023txgnn,
  title={A foundation model for clinician-centered drug repurposing},
  author={Huang, Kexin and others},
  journal={Nature Medicine},
  year={2023}
}