Methodology

From AI prediction to evidence grading — every candidate has a traceable basis for its rating.

Overall pipeline

The platform uses a four-stage pipeline: the TxGNN knowledge graph model predicts potential drug–disease associations, evidence is then collected automatically for each predicted pair, the evidence is graded from L1 to L5, and finally a decision recommendation is issued.

  1. TxGNN prediction: drug–disease relationships predicted with a knowledge graph combined with graph neural networks.
  2. Evidence collection: for each predicted pair, evidence is gathered from ClinicalTrials.gov, PubMed, DrugBank and CDSCO.
  3. Evidence grading: graded L1 to L5, where L1 is strongest (multiple Phase 3 RCTs) and L5 is model prediction only.
  4. Decision recommendation: Go, Proceed, Consider, Explore or Hold, based on the evidence level.

Evidence grading criteria

LevelDefinitionClinical meaning
L1Multiple Phase 3 RCTs / systematic reviewsStrong support; clinical use may be considered
L2Single RCT or multiple Phase 2 trialsModerate support; validation trials can be designed
L3Observational studies / large case seriesPreliminary support; needs further validation
L4Preclinical / mechanistic studiesTheoretical support; far from clinical use
L5Model prediction onlyHypothesis stage; no human evidence yet

Dual-engine prediction

Two methods run in parallel, and a confidence label records whether they agree:

Method Speed Precision Description
Knowledge graph (KG) Fast Lower Inference over DrugBank relations and graph structure
Deep learning (DL) Slow Higher TxGNN graph neural network model
Confidence Source Meaning
very_high KG + DL Both methods agree
high DL only High-scoring deep learning support
medium KG only Knowledge graph support

Regulatory data integration

Drug approval data for India comes from CDSCO. Ingredient names are mapped to the DrugBank vocabulary; ingredients that cannot be mapped — herbal extracts, vaccines, excipients and others not catalogued by DrugBank — are excluded from prediction.


Limitations

  1. Predictions are statistical associations and do not imply causation or clinical efficacy.
  2. An L5 rating means model prediction only, with no supporting human evidence.
  3. Evidence collection depends on public databases; unpublished or unindexed studies are not captured.
  4. Ingredient mapping may miss items because of naming differences.

About the Developer

This platform is developed and operated by 藥提醒科技有限公司 (yao.care, company registration number 83620786, 12F, No. 220, Sec. 2, Taiwan Blvd., West Dist., Taichung City, Taiwan).

InTxGNN is the India site of the company’s “TxGNN Drug Repurposing” product line. The same system is deployed across 30 countries and regions, each named {CC}TxGNN (JpTxGNN, UsTxGNN, DETxGNN, and so on) at {cc}txgnn.yao.care. Product overview: https://www.yao.care/medical/txgnn/.

The TxGNN model itself was developed by the Zitnik Lab at Harvard Medical School and published in Nature Medicine. This platform is the production system 藥提醒科技有限公司 built on top of that model, covering national drug-registration data integration, dual knowledge-graph and deep-learning prediction, PubMed / ClinicalTrials evidence grading, and SMART on FHIR electronic health record integration.


Disclaimer
This report is for academic research reference only and does not constitute medical advice. Always follow your physician's instructions; never adjust medication on your own. Any drug repurposing decision requires full clinical validation and regulatory review.

Reviewed by: 藥提醒科技有限公司 (yao.care)

Copyright © 2026 藥提醒科技有限公司 (yao.care). For research purposes only. Not medical advice.

This site uses Just the Docs, a documentation theme for Jekyll.