quant-phMar 23, 2026

Latent Style-based Quantum Wasserstein GAN for Drug Design

Authors: Julien Baglio, Yacine Haddad, Richard Polifka

Organizations: Center for Quantum Computing and Quantum Coherence (QC2), Department of Physics, University of Basel, Klingelbergstrasse 82, CH-4056 Basel, Switzerland · QuantumBasel, Schorenweg 44b, CH-4144 Arlesheim, Switzerland · Laboratory for High Energy Physics (LHEP), Physics Institute, Faculty of Sciences, University of Bern, Sidlerstrasse 5, CH-3012 Bern, Switzerland · Open Quantum Institute and Quantum Technology Initiative, CERN, Esplanade des Particules 1, CH-1211 Geneva 23, Switzerland

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

The development of new drugs is a tedious, time-consuming, and expensive process, for which the average costs are estimated to be up to around $2.5 billion. The first step in this long process is the design of the new drug, for which de novo drug design, assisted by artificial intelligence, has blossomed in recent years and revolutionized the field. In particular, generative artificial intelligence has delivered promising results in drug discovery and development, reducing costs and the time to solution. However, classical generative models, such as generative adversarial networks (GANs), are difficult to train due to barren plateaus and prone to mode collapse. Quantum computing may be an avenue to overcome these issues and provide models with fewer parameters, thereby enhancing the generalizability of GANs. We propose a new style-based quantum GAN (QGAN) architecture for drug design that implements noise encoding at every rotational gate of the circuit and a gradient penalty in the loss function to mitigate mode collapse. Our pipeline employs a variational autoencoder to represent the molecular structure in a latent space, which is then used as input to our QGAN. Our baseline model runs on up to 15 qubits to validate our architecture on quantum simulators, and a 156-qubit IBM Heron quantum computer in the five-qubit setup is used for inference to investigate the effects of using real quantum hardware on the analysis. We benchmark our results against classical models as provided by the MOSES benchmark suite.

Figures & tables

Figure 1: Schematic representation of the quantum-augmented drug-design pipeline. The model comprises a pre-trained VAE on the MOSES dataset using SMILES input (in purple), that embeds the original molecules into a low-dimensional latent space, and a quantum GAN with a quantum generator (in red) and a classical discriminator (in blue). The features extracted by the VAE from the input dataset are used for the training of the GAN. The final generated SMILES molecules are reconstructed by inversely transforming the features generated by the quantum generator using the pre-trained decoder of the VAE, and finally validated using RDKit.
Figure 2: Styled simple circuit with five qubits and one layer.
Figure 3: Styled basic entangled layer (BEL) circuit with five qubits and one layer. In addition there is a final set of one-qubit RYR_{Y} gates on each qubit to rotate all qubits in the computational basis before performing the final measurement operation.
MetricsNominal VAETuned VAE
Learning rate (lr)3×10−43\times 10^{-4}Scheduled 10−3→10−510^{-3}\rightarrow 10^{-5}
Encoder nln_{l}11
Encoder DhD_{h}256512
Decoder nln_{l}34
Decoder DhD_{h}512512
Decoder dropout00.75
Table 1: VAE hyperparameter settings comparison between nominal (MOSES default) and tuned settings.
MetricsTrain setNominal VAETuned VAE
IntDiv0.8980.8960.896
W(LogP)\mathcal{W}(\text{LogP})0.0910.0870.079
W(SA)\mathcal{W}(\text{SA})0.2680.2750.251
W(QED)\mathcal{W}(\text{QED})0.0270.0250.022
W(Weight)\mathcal{W}(\text{Weight})42.843.941.8
NparamsN_{params}–4,271,2506,472,082
Table 2: Metrics derived from the training set and sets generated by the nominal and tuned VAE. Internal diversity is computed as the mean over the respective samples, whereas the remaining four metrics are the average Wasserstein distances between the sample in question and the test sample for the corresponding metric.
Figure 4: Distributions of metrics derived from the input train set and the generated sets using the nominal and tuned VAE decoder.
Figure 5: Comparison of the loss functions for the nominal (left) and tuned (right) GANs. Each model was run with five different seeds, the shaded band denotes a ±1σ\pm 1\sigma deviation.
MetricsTrain setTuned VAENominal GANTuned GAN
ϵd\epsilon_{d}0.9820.605 ±\pm 0.0170.510 ±\pm 0.0070.481 ±\pm 0.030
ϵv\epsilon_{v}1.0000.868 ±\pm 0.0060.911 ±\pm 0.0060.917 ±\pm 0.012
ϵu\epsilon_{u}1.0000.993 ±\pm 0.0010.986 ±\pm 0.0020.986 ±\pm 0.003
Novelty0.0000.461 ±\pm 0.0260.501 ±\pm 0.0150.536 ±\pm 0.015
Filters0.7340.728 ±\pm 0.0020.723 ±\pm 0.0030.721 ±\pm 0.011
IntDiv0.8980.896 ±\pm 0.0000.902 ±\pm 0.0020.898 ±\pm 0.004
Table 3: Metrics results for the VAE and classical GANs. All GANs use the tuned VAE for the latent space representation. Unlike the other cases, the nominal GAN uses the nominal GAN hyperparameter settings as discussed in the text. The quoted standard deviation comes from running each scenario five times with different seeds for the initialization of the trainable parameters.
Figure 6: Comparison of the metrics for the models used in the analysis. For the VAE, the tuned version is used, for the GAN, not only the tuned VAE is used but also the GAN is tuned itself. Each model was run with five different seeds, the shaded band is defined as ±1σ\pm 1\sigma deviation.
Figure 7: Comparison of the loss functions for the nominal (left) and tuned (right) quantum GANs with the BEL circuit. Each model was run with five different seeds, the shaded band is defined as ±1σ\pm 1\sigma deviation.
MetricsTuned GANSimple QGANBEL QGAN
NparamsN_{params}705,16220110
ϵd\epsilon_{d}0.481±0.0310.481\pm 0.0310.529±0.0030.529\pm 0.0030.523±0.0030.523\pm 0.003
ϵv\epsilon_{v}0.917±0.0120.917\pm 0.0120.831±0.0020.831\pm 0.0020.875±0.0030.875\pm 0.003
ϵu\epsilon_{u}0.986±0.0030.986\pm 0.0030.994±0.0000.994\pm 0.0000.993±0.0010.993\pm 0.001
Novelty0.536±0.0150.536\pm 0.0150.572±0.0030.572\pm 0.0030.548±0.0030.548\pm 0.003
IntDiv0.898±0.0040.898\pm 0.0040.882±0.0000.882\pm 0.0000.883±0.0000.883\pm 0.000
Table 4: Comparison of the performance of the two types of quantum GANs with respect to the classical GAN. All models use tuned VAE and GAN training parameters. The quantum circuit use five qubits, two layers and dual readout to produce a latent vector of dimension 10.
Figure 8: Comparison of the metrics for the two quantum circuits and the classical GAN. All models use tuned VAE and GAN training parameters. The quantum circuit uses five qubits, two layers and dual readout to produce a latent vector of dimension 10. Each model was run with five different seeds and the shaded band is defined as ±1σ\pm 1\sigma deviation.
Figure 9: Distribution of latent vectors for the train, test, classical GAN and BEL QGAN samples.
Latent DimensionGAN  Simple QGANBEL QGAN
( NparamsN_{params} )Readoutnqbn_{qb}nl=2n_{l}=2nl=4n_{l}=4nl=2n_{l}=2nl=4n_{l}=4nl=6n_{l}=6
( nparn_{par} )( nparn_{par} )( nparn_{par} )( nparn_{par} )( nparn_{par} )
10ReferenceDual5-1.00-1.01-0.26-0.27 
(20)(40)(110)(210) 
(705,162)Single10-0.84-0.83-0.10-0.08-0.11
(40)(80)(220)(420)(620)
Table 5: Average ⟨Z0⟩\langle Z_{0}\rangle comparison of different scenarios between classical GANs and QGANs using different setup for the QGANs. The numbers in parenthesis in the first column are the number of trainable parameters and reflect the model capacity (i.e. complexity).
MetricsClassical GANQGAN (noiseless)QGAN (ibm_kingston)
ϵd\epsilon_{d}0.6890.8920.892
ϵv\epsilon_{v}0.9470.9340.955
ϵu\epsilon_{u}0.9901.0001.000
Novelty0.3100.3380.336
IntDiv0.9150.8810.890
Filters0.7190.7060.729
Table 6: Metrics comparison calculated with 2,500 generated molecules, between classical GAN, the style QGAN on noiseless simulator, and the style QGAN inference on ibm_kingston. The quantum generator uses nqb=5n_{qb}=5 , a number of layers nl=2n_{l}=2 , and dual readout.
Figure 10: Distributions of LogP, SA, and QED metrics for the generation of 2,500 molecule using a classical GAN, the style QGAN ran noiseless simulator, and the style QGAN ran on ibm_kingston. The quantum generator uses nqb=5n_{qb}=5 , a number of layers nl=2n_{l}=2 , and dual readout.
Figure 11: Random selection of the 2,500 generated molecules sampled on ibm_kingston, passing the RDKit filters and all new compared to both the training and test dataset. The purple cluster contains molecules with LogP values compatible with injectable drugs; the pink cluster contains molecules with LogP values ideal for oral ingestion; the green cluster contains all other molecules with LogP not in the previous categories and still viable drug candidates with ∣LogP∣<5|\text{LogP}|<5 .
Appendix figures & tables29 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 12: Comparison of metrics distributions between the input train and test sets.
MetricsTrain setTest set
IntDiv0.8980.889
W(LogP)\mathcal{W}(\text{LogP})0.091–
W(SA)\mathcal{W}(\text{SA})0.268–
W(QED)\mathcal{W}(\text{QED})0.027–
W(Weight)\mathcal{W}(\text{Weight})42.8–
Appendix
Table 7: Comparison of metrics between the input train and test sets.
Figure 13: VAE loss function as a function of number of training epochs for the nominal VAE.
NepN_{ep}1002508001,000
GPU runtime9m 10s17m 24s48m 49s52m 35s
IntDiv0.8920.9190.8960.896
W(LogP)\mathcal{W}(\text{LogP})0.1900.0300.0480.048
W(SA)\mathcal{W}(\text{SA})0.2580.26460.27960.2751
W(QED)\mathcal{W}(\text{QED})0.0840.0900.0960.097
W(Weight)\mathcal{W}(\text{Weight})67.651.243.843.9
Appendix
Table 8: VAE metrics comparison for the four NepN_{ep} points. The Wasserstein distance is measured between the test and generated samples. While the internal diversity is maximized, all other metrics are minimized. We also report the GPU runtime for comparison.
Figure 14: Metrics distributions for models trained with different number of epochs. Stabilization of the results starts with 800 epochs.
NgeneratedN_{generated}30,000300,0001,000,000
GPU runtime52m1h 37m2h 12m
Sampling runtime10m55m1h 30m
ϵd\epsilon_{d}0.4790.1680.120
ϵv\epsilon_{v}0.8890.7040.609
ϵu\epsilon_{u}0.9960.9820.977
NCanonicalN_{Canonical}12,72534,76371,118
Appendix
Table 9: Metrics comparison for the nominal VAE training for three different sizes of sets of generated molecules. We also report runtime on the GPU.
lr max3×10−43\times 10^{-4}1×10−41\times 10^{-4}1×10−21\times 10^{-2}5×10−55\times 10^{-5}5×10−35\times 10^{-3}1×10−31\times 10^{-3}
lr min3×10−43\times 10^{-4}1×10−61\times 10^{-6}1×10−41\times 10^{-4}5×10−55\times 10^{-5}5×10−35\times 10^{-3}1×10−51\times 10^{-5}
ϵd\epsilon_{d}0.4790.8050.9380.8420.9360.417
ϵv\epsilon_{v}0.8890.6600.0180.6260.0410.957
ϵu\epsilon_{u}0.9960.9860.9560.9850.9400.997
Novelty0.2370.7020.9250.7680.8560.119
IntDiv0.8960.8930.8870.8910.9290.898
Appendix
Table 10: Comparisons of the results of VAE learning rate tuning, the first column corresponds to the nominal scenario. Individual scenarios are defined by the learning rate minima/maxima at the top of the Table.
Figure 15: Metrics distributions for models trained with different learning rates. Clear breakdown for the 0.01-0.0001 and 0.005 scenarios is visible.
nln_{l}1211
DhD_{h}256256128512
Dropout0.00.50.00.0
ϵd\epsilon_{d}0.4790.5480.4750.5321
ϵv\epsilon_{v}0.8890.8380.8880.8552
ϵu\epsilon_{u}0.9960.9940.9950.9933
Novelty0.2370.3490.2380.327
Appendix
Table 11: Comparisons of the results of VAE encoder parameter tuning, the first column corresponds to the nominal scenario. The scenario variation is highlighted at the top of the table. The dropout is active only for a scenario with more than one encoder layer, in such case value of 0.5 was used.
Figure 16: Metrics distributions for models trained with different encoder settings. Results are very similar between the different scenarios.
nln_{l}3243333
DhD_{h}5125125122561024512512
Dropout0.00.00.00.00.00.250.75
ϵd\epsilon_{d}0.4790.5900.4870.7180.4160.4800.556
ϵv\epsilon_{v}0.8890.8180.8690.7180.9440.9110.875
ϵv\epsilon_{v}0.9960.9900.9960.9880.9970.9960.994
Novelty0.2370.4290.2400.5730.1260.2480.378
Appendix
Table 12: Comparisons of the results of VAE decoder parameter tuning, the first column corresponds to the nominal scenario. The scenario variation is highlighted at the top of the table. The default decoder dropout is zero regardless the number of layers.
Figure 17: Metrics distributions for models trained with different decoder settings. Results are very similar for the different scenarios.
Figure 18: Distributions of 10 latent vectors used as inputs of the VAE decoder.
Figure 19: Metric distribution obtained from SMILES sets which were generated by the VAE decoder from various latent vector distributions. Additional scenario – VAE(gen) – corresponds to a set which is generated by the standard MOSES sampling procedure. The distributions are normalized to unit area.
Figure 20: Correlations between individual latent vectors for the training set, test set, classical GAN training output, and BEL QGAN training output, all for a latent dimension of 10.
Latent dimensionQGANTableFigures
nqbn_{qb}nln_{l}Readout
1052Dual148
1054Dual1521
10102Single1622
10104, 6Single1723
20102Dual1824
Appendix
Table 13: Summary of the different scenarios and links to the corresponding individual tables and figures.
MetricsClassical tunedStyled simple QGANZ0Z_{0}styled BEL QGANZ0Z_{0}
NparamsN_{params}705,16220 110 
ϵd\epsilon_{d}0.481±0.0310.481\pm 0.0310.529±0.0030.529\pm 0.003+1.56+1.560.523±0.0030.523\pm 0.003+1.36+1.36
ϵv\epsilon_{v}0.917±0.0120.917\pm 0.0120.831±0.0020.831\pm 0.002−7.21-7.210.875±0.0030.875\pm 0.003−3.43-3.43
ϵu\epsilon_{u}0.986±0.0030.986\pm 0.0030.994±0.0000.994\pm 0.000+3.50+3.500.993±0.0010.993\pm 0.001+2.90+2.90
Novelty0.536±0.0150.536\pm 0.0150.572±0.0030.572\pm 0.003+2.40+2.400.548±0.0030.548\pm 0.003+0.82+0.82
IntDiv0.898±0.0040.898\pm 0.0040.882±0.0000.882\pm 0.000−4.60-4.600.883±0.0000.883\pm 0.000−4.21-4.21
Appendix
Table 14: Tuned classical GAN vs QGANs with nqb=5n_{qb}=5 , nl=2n_{l}=2 , and dual readout.
MetricsClassical tunedStyled simple QGANZ0Z_{0}Styled BEL QGANZ0Z_{0}
NparamsN_{params}705,16240 210 
ϵd\epsilon_{d}0.481±0.0310.481\pm 0.0310.526±0.0030.526\pm 0.003+1.46+1.460.522±0.0040.522\pm 0.004+1.32+1.32
ϵv\epsilon_{v}0.917±0.0120.917\pm 0.0120.830±0.0020.830\pm 0.002−7.24-7.240.876±0.0040.876\pm 0.004−3.36-3.36
ϵu\epsilon_{u}0.986±0.0030.986\pm 0.0030.994±0.0000.994\pm 0.000+3.42+3.420.993±0.0010.993\pm 0.001+2.80+2.80
Novelty0.536±0.0150.536\pm 0.0150.577±0.0030.577\pm 0.003+2.72+2.720.548±0.0040.548\pm 0.004+0.76+0.76
IntDiv0.898±0.0040.898\pm 0.0040.882±0.0000.882\pm 0.000−4.66-4.660.883±0.0000.883\pm 0.000−4.21-4.21
Appendix
Table 15: Tuned classical GAN vs QGANs with nqb=5n_{qb}=5 , nl=4n_{l}=4 , and dual readout.
Figure 21: Metrics distributions for classical and quantum models. All GANs are trained with a latent dimension of 10. The quantum GANs use nqb=5n_{qb}=5 , nl=4n_{l}=4 , and dual readout.
MetricsClassical tunedStyled simple QGANZ0Z_{0}Styled BEL QGANZ0Z_{0}
NparamsN_{params}716,69240 220 
ϵd\epsilon_{d}0.481±0.0310.481\pm 0.0310.514±0.0040.514\pm 0.004+1.06+1.060.514±0.0020.514\pm 0.002+1.06+1.06
ϵv\epsilon_{v}0.917±0.0120.917\pm 0.0120.833±0.0030.833\pm 0.003−6.99-6.990.887±0.0030.887\pm 0.003−2.45-2.45
ϵu\epsilon_{u}0.986±0.0030.986\pm 0.0030.994±0.0000.994\pm 0.000+3.27+3.270.992±0.0010.992\pm 0.001+2.45+2.45
Novelty0.536±0.0150.536\pm 0.0150.586±0.0030.586\pm 0.003+3.34+3.340.548±0.0020.548\pm 0.002+0.82+0.82
IntDiv0.898±0.0040.898\pm 0.0040.882±0.0000.882\pm 0.000−4.68-4.680.884±0.0000.884\pm 0.000−3.99-3.99
Appendix
Table 16: Tuned classical GAN vs QGANs with nqb=10n_{qb}=10 , nl=2n_{l}=2 , and single readout.
Figure 22: Metrics distributions for classical and quantum models. All GANs are trained with a latent dimension of 10. The quantum GANs use nqb=10n_{qb}=10 , nl=2n_{l}=2 , and single readout.
MetricsClassical tunedStyled simple QGANZ0Z_{0}Styled BEL QGANZ0Z_{0}styled BELZ0Z_{0}
nln_{l}-4 4 6 
NparamsN_{params}705,16280 420 620 
ϵd\epsilon_{d}0.481±0.0310.481\pm 0.0310.515±0.0050.515\pm 0.005+1.08+1.080.516±0.0050.516\pm 0.005+1.13+1.130.515±0.0030.515\pm 0.003+1.10+1.10
ϵv\epsilon_{v}0.917±0.0120.917\pm 0.0120.833±0.0020.833\pm 0.002−6.95-6.950.887±0.0030.887\pm 0.003−2.42-2.420.888±0.0020.888\pm 0.002−2.41-2.41
ϵu\epsilon_{u}0.986±0.0030.986\pm 0.0030.994±0.0000.994\pm 0.000+3.26+3.260.992±0.0010.992\pm 0.001+2.48+2.480.992±0.0010.992\pm 0.001+2.62+2.62
Novelty0.536±0.0150.536\pm 0.0150.586±0.0020.586\pm 0.002+3.37+3.370.549±0.0020.549\pm 0.002+0.87+0.870.550±0.0010.550\pm 0.001+0.93+0.93
Appendix
Table 17: Tuned classical GAN vs QGANs with nqb=10n_{qb}=10 , nl=4,6n_{l}=4,6 , and single readout.
Figure 23: Metrics distributions for classical and quantum models. All GANs are trained with a latent dimension of 10. The quantum GANs use nqb=10n_{qb}=10 , nl=4n_{l}=4 and 6, and single readout.
MetricsClassical tunedStyled simple QGANZ0Z_{0}Styled BEL QGANZ0Z_{0}
Latent dimension2020 20 
NparamsN_{params}716,69240 220 
ϵd\epsilon_{d}0.526 ±\pm 0.0220.507 ±\pm 0.003−0.85-0.850.519 ±\pm 0.005−0.32-0.32
ϵv\epsilon_{v}0.901 ±\pm 0.0080.911 ±\pm 0.004+1.23+1.230.910 ±\pm 0.002+1.22+1.22
ϵu\epsilon_{u}0.990 ±\pm 0.0020.991 ±\pm 0.001+0.25+0.250.991 ±\pm 0.001+0.05+0.05
Novelty0.488 ±\pm 0.0180.484 ±\pm 0.001−0.22-0.220.484 ±\pm 0.002−0.25-0.25
Appendix
Table 18: Quantum GANs with latent dimension of 20, nqb=10n_{qb}=10 , nl=2n_{l}=2 , and dual readout.
Figure 24: Metrics distributions for classical and quantum models. All GANs are trained with a latent dimension of 20. The quantum GANs use nqb=10n_{qb}=10 , nl=2n_{l}=2 , and dual readout.
MetricsClassical tunedStyled simple QGANZ0Z_{0}Styled BEL QGANZ0Z_{0}
Latent dimension3030 30 
NparamsN_{params}728,22260 330 
ϵd\epsilon_{d}0.500 ±\pm 0.0660.496 ±\pm 0.003−0.06-0.060.504 ±\pm 0.002+0.07+0.07
ϵv\epsilon_{v}0.884 ±\pm 0.0320.895 ±\pm 0.002+0.32+0.320.903 ±\pm 0.003+0.57+0.57
ϵu\epsilon_{u}0.990 ±\pm 0.0040.992 ±\pm 0.001+0.42+0.420.991 ±\pm 0.001+0.39+0.39
Novelty0.504 ±\pm 0.0180.476 ±\pm 0.003−1.55-1.550.471 ±\pm 0.003−1.81-1.81
Appendix
Table 19: Quantum GANs with a latent dimension of 30, nqb=15n_{qb}=15 , nl=2n_{l}=2 , and dual readout.
Figure 25: Metrics distributions for classical and quantum models. All GANs are trained with a latent dimension of 30. The quantum GANs use nqb=15n_{qb}=15 , nl=2n_{l}=2 , and dual readout.
Latent Dimension1020Z0Z_{0}30Z0Z_{0}
NparamsN_{params}705,162705,162716,692716,692 728,222728,222 
ϵd\epsilon_{d}0.481±0.0310.481\pm 0.0310.526±0.0220.526\pm 0.022+1.19+1.190.500±0.0660.500\pm 0.066+0.25+0.25
ϵv\epsilon_{v}0.917±0.0120.917\pm 0.0120.901±0.0080.901\pm 0.008−1.16-1.160.884±0.0320.884\pm 0.032−0.96-0.96
ϵu\epsilon_{u}0.986±0.0030.986\pm 0.0030.990±0.0020.990\pm 0.002+1.59+1.590.990±0.0040.990\pm 0.004+0.84+0.84
Novelty0.536±0.0150.536\pm 0.0150.488±0.0180.488\pm 0.018−2.04-2.040.504±0.0180.504\pm 0.018−1.39-1.39
IntDiv0.898±0.0040.898\pm 0.0040.893±0.0020.893\pm 0.002−1.19-1.190.896±0.0030.896\pm 0.003−0.43-0.43
Appendix
Table 20: Classical GAN comparison of different latent dimensions.
Figure 26: Metrics distributions for classical models trained with different latent dimensions.
More in this topic

Molecular Generation

Molecular generation develops computational methods to create chemical structures and conformations for discovery, optimization, and structure elucidation.

See all 234 papers

Explore similar work

CardsList
Feb 17, 2026

Molecular Design beyond Training Data with Novel Extended Objective Functionals of Generative AI Models Driven by Quantum Annealing Computer

Deep generative modeling to stochastically design small molecules is an emerging technology for accelerating drug discovery and development. However, one major issue in molecular generative models is their lower frequency of drug-like compounds. To resolve this problem, we developed a novel framework for optimization of deep generative models integrated with a D-Wave quantum annealing computer, where our Neural Hash Function (NHF) presented herein is used both as the regularization and binarization schemes simultaneously, of which the latter is for transformation between continuous and discrete signals of the classical and quantum neural networks, respectively, in the error evaluation (i.e., objective) function. The compounds generated via the quantum-annealing generative models exhibited higher quality in both validity and drug-likeness than those generated via the fully-classical models, and was further indicated to exceed even the training data in terms of drug-likeness features, without any restraints and conditions to deliberately induce such an optimization. These results indicated an advantage of quantum annealing to aim at a stochastic generator integrated with our novel neural network architectures, for the extended performance of feature space sampling and extraction of characteristic features in drug design.

Hayato Kunugi, Mohsen Rahmani, Yosuke Iyama +8

Inovation to Implementation Laboratories, Central Pharmaceutical Research Institute, Japan Tobacco Inc., 1-1 Murasaki-cho, Takatsuki, 569-1125, Japan. · D-Wave Systems Inc., 3033 Beta Ave, Burnaby, V5G 4M9, Canada.

Molecular GenerationQuantum Generative ModelingDrug DiscoveryMolecular OptimizationHybrid Quantum-Classical MLQuantum Annealing
AskPDF
Feb 27, 2026

Scaling Quantum Machine Learning without Tricks: Full-Resolution and Diverse Image Generation

Quantum generative modeling is a rapidly evolving discipline at the intersection of quantum computing and machine learning. Contemporary quantum machine learning is generally limited to toy examples or heavily restricted datasets with few elements. This is not only due to the current limitations of available quantum hardware but also due to the absence of inductive biases arising from application-agnostic designs. Current quantum solutions must resort to tricks to scale down high-resolution images, such as relying heavily on dimensionality reduction or utilizing multiple quantum models for low-resolution image patches. Building on recent developments in classical image loading to quantum computers, we circumvent these limitations and train quantum Wasserstein GANs on the established classical MNIST and Fashion-MNIST datasets. Using the complete datasets, our system generates full-resolution images across all ten classes and establishes a new state-of-the-art performance with a single end-to-end quantum generator without tricks. As a proof-of-principle, we also demonstrate that our approach can be extended to color images, exemplified on the Street View House Numbers dataset. We analyze how the choice of variational circuit architecture introduces inductive biases, which crucially unlock this performance. Furthermore, enhanced noise input techniques enable highly diverse image generation while maintaining quality. Finally, we show promising results even under quantum shot noise conditions.

Jonas Jäger, Florian J. Kiwit, Carlos A. Riofrío

BMW Group, Munich, Germany · Department of Computer Science and Institute of Applied Mathematics, University of British Columbia (UBC), Vancouver, Canada · Stewart Blusson Quantum Matter Institute (QMI), Vancouver, Canada +1

Quantum Generative ModelingImage GenerationQuantum Machine LearningVariational Quantum CircuitsGenerative Adversarial Networks
AskPDF
Jun 11, 2026

An LLM System for Autonomous Variational Quantum Circuit Design

The design of high performing quantum circuits remains largely dependent on human expertise. We introduce an autonomous agentic framework that employs large language models (LLMs) to conduct iterative quantum circuit designs under explicit design constraints. Our system integrates seven components: Exploration, Generation, Discussion, Validation, Storage, Evaluation, and Review. These components form a closed-loop workflow that combines web-based knowledge acquisition, literature-grounded critique, executable code generation, and experimental feedback. We evaluate the framework on two tasks: quantum feature map construction for quantum machine learning and ansatz generation for variational quantum eigensolver applications in quantum chemistry. In image classification benchmarks, the best generated feature map outperforms representative quantum feature maps and, when scaled to larger qubit counts, surpasses the classical radial basis function kernel. In molecular ground state estimation across seven molecules, the generated ansatz attains competitive accuracy with widely used chemically inspired and hardware-efficient constructions while satisfying the imposed scaling constraints. These results establish LLM driven agentic system as a viable paradigm for automated quantum circuit design and illustrate how AI systems can participate in iterative scientific optimization workflows across scientific domains.

Kenya Sakka, Wataru Mizukami, Kosuke Mitarai

Center for Quantum Information and Quantum Biology, The University of Osaka, 1-2 Machikaneyama, Toyonaka 560-0043, Japan · Graduate School of Engineering Science, The University of Osaka 1-3 Machikaneyama, Toyonaka, Osaka 560-8531, Japan

Quantum ChemistryQuantum Kernel MethodsVariational Quantum CircuitsLarge Language Model-Guided Scientific Discovery
AskPDF