Work

Improved normalization in quantum-classical hybrid models for anti-cancer drug response prediction

Year: 2026
Type: preprint
Abstract: Quantum-classical Hybrid Machine Learning (QHML) models are AI systems that combine elements of classical algorithms, like neural networks, and quantum circuits to perform supervised inference tasks. They are recognized for their ability to perform comparably to classical models while reducing the number of trainable parameters. These qualities may offer unique advantages in a scenario of anti-cancer drug response prediction, where the number of available samples is typically small. However, we observed that QHML models appear to be very sensitive to the data encoding used at the interface between the neural networks and the quantum circuits, with suboptimal choices leading to stability issues. To address this problem, we propose a novel strategy that uses a hyperbolic tangent function with moderated gradient to normalize quantum circuit inputs. This method transforms the outputs of the neural networks without concentrating them at the extreme ends of the value range. The proposed normalization gave the best test AUC for three of the four tested anti-cancer drugs from the GDSC panel and we successfully confirmed that simulator-based performance for docetaxel response prediction (AUC$\simeq$0.87) can be reproduced on a 20-qubit trapped-ion quantum computer. The results confirmed that the introduction of quantum circuits can degrade performance unless the normalization is chosen appropriately. This study emphasizes the importance of designing normalization functions to fill the gap between quantum and classical parts in QHML.
Language: English
Open Access status: green
OpenAlex ID: W7233955858
Cites: 34
Cited by:
Related to: 10