Hybrid quantum cycle generative adversarial network for small molecule generation

TL;DR

This work tackles the high cost of drug design by deploying hybrid quantum-classical generative models for small-molecule graphs. It introduces HQ-MolGAN, HQ-Cycle-MolGAN, and Hybrid Quantum Cycle MolGAN, integrating parametrized quantum circuits as the generator and a Cycle component to stabilize training, evaluated on QM9 and PC9. The results show up to $30\%$ improvement in the QED score and improved SA and LogP metrics, with HQ-MolGAN outperforming classical MolGAN and the Cycle variants boosting stability and sample diversity; the quantum-cycle models demonstrate resilience to noise in IBM Brisbane simulations. The findings underscore the potential of near-term quantum-enhanced architectures in accelerating drug discovery and guide future work on quantum-classical balance and dataset expansion for broader chemical spaces.

Abstract

The drug design process currently requires considerable time and resources to develop each new compound that enters the market. This work develops an application of hybrid quantum generative models based on the integration of parametrized quantum circuits into known molecular generative adversarial networks, and proposes quantum cycle architectures that improve model performance and stability during training. Through extensive experimentation on benchmark drug design datasets, QM9 and PC9, the introduced models are shown to outperform the previously achieved scores. Most prominently, the new scores indicate an increase of up to 30% in the quantitative estimation of druglikeness. The new hybrid quantum machine learning algorithms, as well as the achieved scores of pharmacokinetic properties, contribute to the development of fast and accurate drug discovery processes.

Figures (8)

• Figure 1: Histograms of distribution of values of QED, SA and LogP scores in QM9 and PC9 datasets. The mean QED and LogP scores for molecules in the PC9 dataset are greater than those in QM9, while the mean score for SA is lower.
• Figure 2: (a) Structure of HQ-Cycle-MolGAN: Generator (G), Discriminator (D), Cycle Component (C). The part highlighted in green is the same as HQ-MolGAN. (b) Illustration of the work of the Cycle Components. Suppose $Z$ is a space of normally distributed noise vectors, and $Y$ is a chemical space of datasets. The Generator maps $Z$ to some chemical space $Y'$, and after the Cycle Component restores vector $G(Z)$ back to noise. The accuracy of these restorations is then optimized. (c) Quantum Depth-Infused Neural Network Layer used as the HQ-Cycle component.
• Figure 3: Chart of (a) QED, (b) LogP, and (c) SA scores during the training of classical MolGAN and Hybrid MolGAN. It can be seen that while MolGAN limits its scores to a narrow beam of values even after $50,000$ iterations, Hybrid MolGAN presents a wider range of compounds, covering greater scores of key metrics.
• Figure 4: (a) Samples generated by HQ-MolGAN-VVRQ trained on QM9. (b) "High entropy state": HQ-MolGAN-VVRQ generated inappropriate samples and RDKit rewarded them with an average metric of LogP $\propto$ 0.9. (c) Samples generated by HQ-Cycle-MolGAN-VVRQ trained on both datasets. (d) Samples generated by MolGAN with HQ-Cycle trained on both datasets.
• Figure 5: Figures (a)-(d) present a comparison of the combined losses during training on the QM9 dataset: (a) MolGAN and Cycle-MolGAN. Cycle-MolGAN has a more stable training process compared to the MolGAN. (b) HQ-Cycle-MolGAN-EFQ and HQ-MolGAN-VVRQ. (c) MolGAN an HQ-Cycle-Component MolGAN. (d) HQ-Cycle-MolGAN-EFQ versus Hybrid-MolGAN-EFQ. No significant impact of the Cycle component on the loss curve is observed. Figures (e)-(f) present charts for the IBM Brisbane execution: (e) Graph of the relative errors of the simulators' probability matrices with respect to the number of shots. (f) Comparison of the probabilities generated by the noisy and ideal simulators using $2\times 10^5$ shots.