SgTxGNN - Singapore: Drug Repurposing

本專案由 藥提醒科技有限公司（yao.care）開發與維運。 SgTxGNN 站台：https://sgtxgnn.yao.care/｜產品總覽：https://www.yao.care/medical/txgnn/

Drug repurposing predictions for Singapore HSA-approved drugs using the TxGNN model.

Disclaimer

• The results of this project are for research purposes only and do not constitute medical advice.
• Drug repurposing candidates require clinical validation before application.

Project Overview

Report Statistics

Item	Count
Drug Reports	966
Total Predictions	10,473,055
Unique Drugs	1,041
Unique Indications	17,184
DDI Data	160,235
FHIR Resources	999 MK / 4,189 CUD

Evidence Level Distribution

Evidence Level	Report Count	Description
L1	11	Multiple Phase 3 RCTs
L2	34	Single RCT or multiple Phase 2
L3	22	Observational studies
L4	54	Preclinical / mechanistic studies
L5	602	Computational prediction only
N/A	243	Known uses only (no new prediction)

By Source

Source	Predictions
DL	10,454,986
KG	14,364
KG + DL	3,705

By Confidence

Confidence	Predictions
very_high	1,515
high	479,332
medium	918,852
low	9,073,356

---

Prediction Methods

Method	Speed	Accuracy	Requirements
Knowledge Graph	Fast (seconds)	Lower	No special requirements
Deep Learning	Slow (hours)	Higher	Conda + PyTorch + DGL

Knowledge Graph Method

uv run python scripts/run_kg_prediction.py

Metric	Value
HSA Total Drugs	7,117
Mapped to DrugBank	6,038 (84.8%)
Repurposing Candidates	18,069

Deep Learning Method

conda activate txgnn
PYTHONPATH=src python -m sgtxgnn.predict.txgnn_model

Metric	Value
Total DL Predictions	2,056,697
Unique Drugs	1,041
Unique Indications	17,184

Score Interpretation

The TxGNN score represents the model's confidence in a drug-disease pair, ranging from 0 to 1.

Threshold	Meaning
>= 0.9	Very high confidence
>= 0.7	High confidence
>= 0.5	Moderate confidence

Score Distribution

Threshold	Meaning
≥ 0.9999	Extremely high confidence, model's most confident predictions
≥ 0.99	Very high confidence, worth prioritizing for validation
≥ 0.9	High confidence
≥ 0.5	Moderate confidence (sigmoid decision boundary)

Evidence Level Definitions

Level	Definition	Clinical Significance
L1	Phase 3 RCT or systematic review	Can support clinical use
L2	Phase 2 RCT	Can consider for use
L3	Phase 1 or observational study	Requires further evaluation
L4	Case report or preclinical research	Not recommended yet
L5	Computational prediction only, no clinical evidence	Requires further research

Important Reminders

1. High scores do not guarantee clinical efficacy: TxGNN scores are knowledge graph-based predictions that require clinical trial validation.
2. Low scores do not mean ineffective: The model may not have learned certain associations.
3. Recommended to use with validation pipeline: Use this project's tools to review clinical trials, literature, and other evidence.

Validation Pipeline

flowchart TD
    A["TxGNN Prediction Results"] --> B
    subgraph B["Step 1: DrugBundle Collector"]
        B1["Drug-level: HSA, DDI, DrugBank"]
        B2["Indication-level: ClinicalTrials, PubMed, ICTRP"]
    end
    B --> |"drug_bundle.json"| C
    subgraph C["Step 2: Evidence Pack Generator"]
        C1["Programmatic data transfer (100%)"]
        C2["+ LLM Analysis (L1-L5)"]
    end
    C --> |"drug_evidence_pack.json/md"| D
    subgraph D["Step 3: Notes Writer"]
        D1["drug_pharmacist_notes.md"]
        D2["drug_sponsor_notes.md"]
    end

---

Quick Start

Step 1: Download Data

File	Download
HSA Data	data.gov.sg
node.csv	Harvard Dataverse
kg.csv	Harvard Dataverse
edges.csv	Harvard Dataverse
model_ckpt.zip	Google Drive

Step 2: Install Dependencies

uv sync

Step 3: Process Drug Data

uv run python scripts/process_fda_data.py

Step 4: Prepare Vocabulary Data

uv run python scripts/prepare_external_data.py

Step 5: Run Knowledge Graph Prediction

uv run python scripts/run_kg_prediction.py

Step 6: Set Up Deep Learning Environment

conda create -n txgnn python=3.11 -y
conda activate txgnn
pip install torch==2.2.2 torchvision==0.17.2
pip install dgl==1.1.3
pip install git+https://github.com/mims-harvard/TxGNN.git
pip install pandas tqdm pyyaml pydantic ogb

Step 7: Run Deep Learning Prediction

conda activate txgnn
PYTHONPATH=src python -m sgtxgnn.predict.txgnn_model

---

Resources

TxGNN Core

• TxGNN Paper - Nature Medicine, 2024
• TxGNN GitHub
• TxGNN Explorer

Data Sources

Category	Data	Source	Note
Drug Data	HSA	data.gov.sg	Singapore
Knowledge Graph	TxGNN KG	Harvard Dataverse	17,080 diseases, 7,957 drugs
Drug Database	DrugBank	DrugBank	Drug ingredient mapping
Drug Interactions	DDInter	DDInter	DDI pairs
Drug Interactions	Guide to PHARMACOLOGY	IUPHAR/BPS	Approved-drug target data (not interactions)
Clinical Trials	ClinicalTrials.gov	CT.gov API v2	Clinical trials registry
Clinical Trials	WHO ICTRP	ICTRP API	International clinical trials platform
Literature	PubMed	NCBI E-utilities	Medical literature search
Name Mapping	RxNorm	RxNav API	Drug name standardization bridge

Model Downloads

File	Download	Note
Pretrained Model	Google Drive	model_ckpt.zip
node.csv	Harvard Dataverse	Node Data
kg.csv	Harvard Dataverse	Knowledge Graph Data
edges.csv	Harvard Dataverse	Edge Data (DL)

Project Introduction

Directory Structure

SgTxGNN/
├── README.md
├── CLAUDE.md
├── pyproject.toml
│
├── config/
│   └── fields.yaml
│
├── data/
│   ├── kg.csv
│   ├── node.csv
│   ├── edges.csv
│   ├── raw/
│   ├── external/
│   ├── processed/
│   │   ├── drug_mapping.csv
│   │   ├── repurposing_candidates.csv
│   │   ├── txgnn_dl_predictions.csv.gz
│   │   └── integration_stats.json
│   ├── bundles/
│   └── collected/
│
├── src/sgtxgnn/
│   ├── data/
│   │   └── loader.py
│   ├── mapping/
│   │   ├── normalizer.py
│   │   ├── drugbank_mapper.py
│   │   └── disease_mapper.py
│   ├── predict/
│   │   ├── repurposing.py
│   │   └── txgnn_model.py
│   ├── collectors/
│   └── paths.py
│
├── scripts/
│   ├── process_fda_data.py
│   ├── prepare_external_data.py
│   ├── run_kg_prediction.py
│   └── integrate_predictions.py
│
├── docs/
│   ├── _drugs/
│   ├── fhir/
│   │   ├── MedicationKnowledge/
│   │   └── ClinicalUseDefinition/
│   └── smart/
│
├── model_ckpt/
└── tests/

Legend: 🔵 Project development | 🟢 Local data | 🟡 TxGNN data | 🟠 Validation pipeline

Data Flow

flowchart TD
    FDA["HSA Data"] --> proc["process_fda_data.py"]
    TxGNN["TxGNN Data"] --> prep["prepare_external_data.py"]

    proc --> json["sg_fda_drugs.json"]
    prep --> ext["data/external/"]

    json --> norm["normalizer.py"]
    ext --> norm

    norm --> drug_map["drug_mapping.csv"]
    drug_map --> ind_map["indication_mapping.csv"]

    ind_map --> KG["Knowledge Graph"]
    ind_map --> DL["Deep Learning"]

    KG --> kg_out["repurposing_candidates.csv"]
    DL --> dl_out["txgnn_dl_predictions.csv"]

---

Citation

If you use this dataset or software, please cite:

@software{sgtxgnn2026,
  author       = {Yao.Care},
  title        = {SgTxGNN: Drug Repurposing Validation Reports for Singapore HSA Drugs},
  year         = 2026,
  publisher    = {GitHub},
  url          = {https://github.com/yao-care/SgTxGNN}
}

Also cite the original TxGNN paper:

@article{huang2023txgnn,
  title={A foundation model for clinician-centered drug repurposing},
  author={Huang, Kexin and Chandak, Payal and Wang, Qianwen and Haber, Shreyas and Zitnik, Marinka},
  journal={Nature Medicine},
  year={2024},
  doi={10.1038/s41591-024-03233-x}
}

維運：資料還原與許可證更新

完整還原

git 只放程式碼與網站頁面；許可證原始資料、證據包、藥師報告原稿、預測結果放在本 repo 的 Release。

git clone https://github.com/yao-care/sgtxgnn.git && cd sgtxgnn
uv sync                                  # Python 相依套件
bash scripts/restore_data.sh             # 取最新的 data-* Release，還原 data/（含校驗）
bash scripts/restore_data.sh --with-kg   # 另加共用知識圖譜與預測模型（只有重跑預測才需要）

需要 GitHub CLI（gh auth login）與 uv。

改頁工具鏈

事實層、checker、許可證狀態計算等放在 tools/txgnn/.claude/，是母 repo yao-care/txgnn 的 .claude/ 複本（版本見 tools/txgnn/VERSION）。 規則要改請改母 repo，再用 python3 .claude/sync_site_tools.py SgTxGNN 同步過來，不要直接改這裡的複本。 在本 repo 單獨使用時，先組工作目錄：

WS=$(bash scripts/txgnn_ws.sh)
cd "$WS"
python3 .claude/dump_license_facts_20261021.py SgTxGNN       # 依證據包算市場狀態（checker 要用）
python3 .claude/checker_trend.py --no-append SgTxGNN           # 全部藥物頁 checker，必須「非 0 0」

許可證排程更新（GitHub Actions）

.github/workflows/license-refresh.yml 每週一 UTC 03:17 自動執行，也可以在 Actions 頁手動觸發（可選只試跑）：

1. 從 Release 還原 data/（scripts/restore_data.sh）
2. 從主管機關下載最新許可證資料（scripts/license_adapter.py），檢查筆數與格式
3. 更新證據包的許可證狀態 → 重算市場狀態 → 同步許可證表 → 事實層改頁 → checker 全量必須 0 → 衍生檔 → gate
4. 全部通過才推上 main 並觸發部署；更新後的 data/ 另存成新的 data-YYYYMMDD Release

以下情況一律不推、自動開 issue（label license-refresh）：下載失敗或格式改變、狀態變動超過兩成、 改到的藥物頁超過 min(200, 5%)、checker／gate／h1 沒過、有人工查核紀錄的頁會被改到。全程不使用 AI。