Predict treatments
on a knowledge graph

One graph.
Two model paths.

1. 01Step

   The problem

   Hetionet is an open biomedical knowledge graph. This project predicts Compound–treats–Disease (CtD) links — which drugs might treat which diseases — from known treatments and graph structure.

2. 02Step

   The approach

   Known treatment pairs are positives; unknown pairs are negatives. Each compound and disease gets an embedding. Classical models (forests, logistic regression) run beside quantum kernel classifiers (QSVC / VQC). A stacking ensemble combines both paths.

3. 03Step

   What to show

   The published target was test PR-AUC above 0.70. The best logged run is a Pauli stacking ensemble at 0.7987, ahead of the tuned classical forests. That comparison — hybrid vs classical baseline — is the lab.

Test PR-AUC
from the log

Figures below are the published experiment log in the public repo. Target was PR-AUC above 0.70. Best run: stacking ensemble (Pauli) at 0.7987.

Ensemble-QC-stacking (Pauli) · Hybrid ensemble

0.7987

RandomForest-Optimized · Classical

0.7838

ExtraTrees-Optimized · Classical

0.7807

Ensemble-QC-stacking (ZZ) · Hybrid ensemble

0.7408

QSVC-Optimized · Quantum

0.7216