Quantum-Inspired Hamiltonian Feature Extraction for ADMET Prediction: A Simulation Study

TL;DR

This paper introduces a quantum-inspired Hamiltonian feature extraction method for ADMET prediction, leveraging quantum simulation to improve molecular property prediction accuracy and interpretability.

Contribution

It presents a novel quantum-inspired encoding of molecular fingerprints into Hamiltonians guided by mutual information, enhancing predictive performance on ADMET benchmarks.

Findings

Achieved state-of-the-art AUROC 0.673 on CYP3A4 prediction.

Improved performance on 8 out of 10 ADMET tasks compared to classical methods.

Quantum-derived features significantly contribute to model importance despite small feature size.

Abstract

Predicting absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties remains a critical bottleneck in drug discovery. While molecular fingerprints effectively capture local structural features, they struggle to represent higher-order correlations among molecular substructures. We present a quantum-inspired feature extraction method that encodes molecular fingerprints into a parameterized Hamiltonian, using mutual information (MI) to guide entanglement structure. By simulating quantum evolution on GPU-accelerated backends, we extract expectation values that capture pairwise and triadic correlations among fingerprint bits. On ten Therapeutic Data Commons (TDC) ADMET benchmarks, our method achieves state-of-the-art performance on CYP3A4 substrate prediction (AUROC 0.673 0.004) and improves over classical baselines on 8/10 tasks. SHAP (SHapley Additive exPlanations) analysis reveals that quantum-derived features contribute up to 33% of model importance despite comprising only 1.6% of features, demonstrating that Hamiltonian encoding concentrates predictive signal. This simulation study establishes the foundation for hardware validation on near-term quantum devices.