🧬 tox21_hybrid_quantum_model

A research pipeline building and validating a multi-task GNN-quantum architecture for toxicity prediction on the Tox21 dataset (12 simultaneous toxicity endpoints), tested against a parameter-matched classical control across multiple random seeds.

📋 Overview

This project builds a graph neural network (GNN) encoder that learns molecular representations directly from atomic and bond structure, feeding into an 8-qubit variational quantum circuit with 12 independent task heads. It is evaluated against a parameter-matched classical control and validated across multiple random seeds — the same statistical discipline expected of published QML benchmarking work.

🏗️ Architecture

1. Stabilized GNN encoder — converts molecular graphs (atom features + adjacency matrix) into continuous latent vectors, replacing fixed Morgan fingerprints with learned structural representations.
2. 8-qubit variational quantum circuit — AngleEmbedding followed by 4 layers of RY/RZ rotations and CNOT entangling gates.
3. Multi-head decoders — 12 isolated classical heads, one per toxicity task, to prevent negative transfer between unrelated assays.
4. Training — weighted BCE loss, masked loss for missing per-task labels, early stopping, gradient clipping, ReduceLROnPlateau.

📊 Results

Multi-task GNN-quantum model (12 tasks, 3-seed validation)

Model	Mean AUC (3 seeds)	Std
8-qubit quantum-hybrid GNN	0.6769	± 0.0204
Parameter-matched classical GNN control	0.6545	± 0.0262

Per-seed breakdown:

Seed	Classical GNN	Quantum GNN
42	0.6492	0.6487
101	0.6254	0.6856
777	0.6889	0.6964

Quantum led on 2 of 3 seeds and trailed narrowly on the third; the mean across seeds favors quantum, but the per-seed spread is exactly why the gap does not clear the variance threshold below.

Optimal decision threshold (G-mean optimized): 0.4066

The result is consistent with a competitive-but-unconfirmed pattern: the quantum-hybrid model scored higher on average across seeds. To check whether that gap reflects a real effect rather than a favorable split, it was tested against:

• A parameter-matched classical control — a classical block sized to match the quantum layer's trainable parameter count, not a bare linear layer, so the comparison isn't quantum vs. an artificially weak baseline.
• Multi-seed validation — the same experiment re-run across 3 random seeds to separate signal from seed-to-seed noise.

At this qubit count and dataset scale, the observed gap (+0.0224 AUC) sits inside the combined variance of both models' seed distributions. This project reports that plainly rather than rounding it up to a proven advantage — the willingness to run and report that check is part of what this repo is meant to demonstrate.

📁 Repo contents

tox21_hybrid_quantum_model.ipynb   # main notebook — full pipeline
README.md
requirements.txt
.gitignore

expert_scaled_multihead_best.pth (trained weights) and production_metadata_log.json (validation logs) are generated by the notebook but excluded from version control by .gitignore — see Setup below to regenerate them locally.

⚙️ Setup

1. Clone the repo
2. pip install -r requirements.txt
3. Place tox21.csv.gz in the same folder as the notebook (download here)
4. Run the notebook in Jupyter or Colab

Running in Colab: the first cell installs rdkit, pennylane, and pennylane-lightning automatically — no local environment setup needed.

📦 Dependencies

torch>=2.0
pennylane>=0.35
pennylane-lightning
rdkit>=2023.03
scikit-learn
numpy
pandas
matplotlib

🔭 Next steps

• Scaffold-split evaluation (holding out entire molecule families) to test generalization beyond a random train/test split
• Additional seeds / a formal paired significance test to sharpen the confidence interval on the quantum-classical gap
• Evaluation on real quantum hardware rather than simulation

📄 License

MIT

Built With

• jupyter-notebook