nature biotechnology

Brief Communication https://doi.org/10.1038/s41587-024-02526-3

Quantum-computing-enhanced algorithm
unveils potential KRAS inhibitors

Received: 7 May 2024 Mohammad Ghazi Vakili1,2, Christoph Gorgulla 3,4 , Jamie Snider5,
AkshatKumar Nigam6 , Dmitry Bezrukov7, Daniel Varoli8, Alex Aliper7,
Accepted: 6 December 2024
Daniil Polykovsky 9, Krishna M. Padmanabha Das10,11, Huel Cox III11,
Published online: xx xx xxxx Anna Lyakisheva5, Ardalan Hosseini Mansob 5,12, Zhong Yao5, Lela Bitar5,13,
Danielle Tahoulas 5,14, Dora Čerina 14,15, Eugene Radchenko 7, Xiao Ding7,
Check for updates
Jinxin Liu7, Fanye Meng7, Feng Ren 7, Yudong Cao16, Igor Stagljar 5,12,14,17 ,
Alán Aspuru-Guzik 1,2,18,19,20,21 & Alex Zhavoronkov 7

We introduce a quantum–classical generative model for small-molecule

design, specifically targeting KRAS inhibitors for cancer therapy. We apply
the method to design, select and synthesize 15 proposed molecules that
could notably engage with KRAS for cancer therapy, with two holding
promise for future development as inhibitors. This work showcases the
potential of quantum computing to generate experimentally validated hits
that compare favorably against classical models.

Drug discovery is a multifaceted and resource-intensive process encom- By merging the advancements in quantum machine learning with
passing the discovery, development and comprehensive testing of new traditional drug discovery, the industry is shifting toward innovative
molecules. Typically extending over a decade and incurring substantial computational strategies12. While purely classical algorithms have
costs, the pharmaceutical industry faces substantial financial risks1–3. made notable strides in drug discovery, hybrid classical–quantum
The pressing need for efficiency and innovation in drug discovery approaches offer unique advantages because of the ability of quan-
has led to integrating advanced computational tools into traditional tum circuit Born machines (QCBMs) to leverage quantum effects such
pharmaceutical research methodologies. Concurrently, generative as superposition and entanglement13,14. The introduction of QCBMs
modeling has emerged as a transformative technology in molecule illustrates this advancement, offering a generative model that can
design4–7. Generative models use machine learning techniques to under- outperform classical ones in certain aspects. QCBMs are quantum
stand the underlying distribution of atoms and bonds in a specified generative models that leverage quantum circuits to learn complex
dataset, which are then used to construct molecules with predefined probability distributions, enabling them to generate new samples
properties, a process known as inverse molecular design8–10. These that resemble the training data15. Hibat-Allah et al.16 demonstrated
models excel at exploring the vast drug-like chemical space (~1060 the superior generalization capabilities of QCBMs, which can pro-
molecules), identifying potential molecules efficiently11. duce cardinality distributions beyond the training set and overcome

1
Department of Computer Science, University of Toronto, Toronto, Ontario, Canada. 2Department of Chemistry, University of Toronto, Toronto, Ontario,
Canada. 3Department of Structural Biology, St. Jude Children’s Research Hospital, Memphis, TN, USA. 4Department of Physics, Harvard University,
Cambridge, MA, USA. 5Donnelly Centre, Temerty Faculty of Medicine, University of Toronto, Toronto, Ontario, Canada. 6Department of Computer
Science, Stanford University, Stanford, CA, USA. 7Insilico Medicine AI Limited, Abu Dhabi, UAE. 8Zapata AI, Boston, MA, USA. 9Insilico Medicine Canada,
Inc., Montreal, Québec, Canada. 10Department of Cancer Biology, Dana-Farber Cancer Institute, Boston, MA, USA. 11Department of Biological Chemistry
and Molecular Pharmacology, Harvard Medical School, Harvard University, Boston, MA, USA. 12Department of Molecular Genetics, University of Toronto,
Toronto, Ontario, Canada. 13Department for Lung Diseases Jordanovac, Clinical Hospital Centre Zagreb, University of Zagreb, Zagreb, Croatia. 14Donnelly
Centre, Department of Biochemistry, Temerty Faculty of Medicine, University of Toronto, Toronto, Ontario, Canada. 15Department of Oncology, University
Hospital Center Split, School of Medicine, University of Split, Split, Croatia. 16AQI, Inc., Boston, MA, USA. 17Mediterranean Institute for Life Sciences
(MedILS), School of Medicine, University of Split, Split, Croatia. 18Department of Chemical Engineering and Applied Chemistry, University of Toronto,
Toronto, Ontario, Canada. 19Department of Materials Science and Engineering, University of Toronto, Toronto, Ontario, Canada. 20Vector Institute for
Artificial Intelligence, Toronto, Ontario, Canada. 21Fellow, Canadian Institute for Advanced Research (CIFAR), Toronto, Ontario, Canada.
e-mail: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]

Nature Biotechnology
Brief Communication https://doi.org/10.1038/s41587-024-02526-3

a Training data generation

Data ity
Hit expansion l divers Virtual screening with
ructura
mining sing st
Increa

100 M
molecules
screened

KRAS structure Ligand library

Mutation sites Similarity > 0.6 Similarity > 0.4

Experimental ligands STONED SELFIES Virtual screening hits

650 molecules 850K molecules Top 250K molecules

Training set
b Generation of new molecules

Generative model

QCBM LSTM Filter

generator generator discriminator

Quantum computer (IBM) Classical computer/cloud Remote API

c KRAS inhibitor design campaign

Compound selection Synthesis of hits Experimental validation

d Identification of inhibitor candidates

Classical samples
1M compounds
Selected 3 compounds
Qunatum samples Selected 8 compounds
1M compounds Selected 4 compounds
Simulated samples
Molecular docking Filtering based on Sorting based Structural Synthetic
1M compounds via Chemistry42 pharmacophores on PLI score novelty complexity

Classical Quantum (HW) Quantum (SIM)

Fig. 1 | Schematic representation of the hybrid quantum–classical framework chemical structures, while the QCBM, trained on the output from the LSTM,
for KRAS ligand development. a, The initial phase involved assembling a generated complex, high-dimensional probability distributions. This workflow
training dataset, starting with 650 experimentally verified KRAS inhibitors incorporated Chemistry42 as a reward function to encourage the production
sourced from the literature. Using the STONED–SELFIES algorithm, analogs of structurally diverse and synthesizable molecules. c, Workflow for KRAS
of these inhibitors were generated, expanding the dataset to approximately inhibitor design, detailing the process from computational compound selection
850,000 compounds. This was further augmented by adding 250,000 top to laboratory synthesis and experimental validation. d, A total of 1 million
candidates from a virtual screening of the REAL ligand library against KRAS, compounds (classical samples from the LSTM, quantum samples from QCBM on
creating a total dataset of over 1 million molecules. b, In the generation phase, quantum hardware and simulated quantum samples on classical hardware) were
the dataset was used to train our generative model, consisting of both a classical evaluated by Chemistry42 to filter out unsuitable candidates and rank the rest by
LSTM network and a QCBM. The LSTM network processed sequential data of their PLI scores. Finally, 15 promising compounds were selected for synthesis.

learning challenges such as barren plateaus. The integration of tensor classical systems and quantum prior size. Additionally, Zeng et al.20
networks furthers their effectiveness17,18. However, the challenges presented a quantum–classical hybrid for image generation, high-
in quantum data processing and circuit trainability have led to the lighting the ability to encode conditions into quantum circuits with
development of hybrid algorithms that harness the strengths of extra qubits.
both quantum and classical machine learning. Manuel et al.19 show- Here, we propose a hybrid quantum–classical model that
cased how a hybrid quantum circuit generative adversarial network addresses qubit limitations and combines quantum and classical
can enhance target space exploration, overcoming limitations of approaches to generate compounds targeting the KRAS protein,

Nature Biotechnology
Brief Communication https://doi.org/10.1038/s41587-024-02526-3

which is known for its intricate complexity and historical resistance complex, high-dimensional probability distributions more efficiently
to drug discovery efforts21,22. As illustrated in Fig. 1, our workflow was than classical models27. Entanglement enables the creation of correla-
structured into three pivotal stages. The first stage was the genera- tions between qubits, capturing intricate dependencies within the prior
tion of training data, beginning with creating a dataset comprising distribution. This capability is particularly advantageous in generative
approxi­mately 650 known KRAS inhibitors from the literature. To models, as it allows for a more accurate representation of the underly-
enhance this dataset, we used VirtualFlow 2.0 to screen 100 million ing distributions in complex datasets, leading to the improved genera-
molecules from the Enamine REAL library, selecting the top 250,000 tion of molecular structures. Additionally, these quantum properties
with the best docking scores23. Additionally, we used the STONED help to escape barren plateaus more effectively during optimization—a
(superfast traversal, optimization, novelty, exploration and discovery) common challenge in machine learning models17. Consequently, this
algorithm on the SELFIES (self-referencing embedded strings) mole­ resulted in a more comprehensive exploration of the solution space,
cular representation of known inhibitors, generating structurally potentially uncovering molecular structures that our naive LSTM could
similar compounds. After applying synthesizability filtering, this not generate within the same time frame.
process added 850,000 molecules to our training set24. Furthermore, Furthermore, our KRAS inhibitor campaign validated com-
we used Chemistry42 to validate the molecules generated in this pounds derived from both the hybrid quantum–classical and vanilla
workflow25. We merged data from various sources to compile a single approaches using the Chemistry42 platform’s structure-based drug
dataset containing 1.1 million data points, all used for training our gen- design workflow. Within our study, 15 compounds were selected after
erative model. The next stage was the generation of new molecules, filtering generated compounds (Fig. 1) using Chemistry42 and synthe-
which combined (1) the QCBM16 using a 16-qubit processor to generate sized for experimental analysis. The two compounds demonstrating
a prior distribution; (2) a long short-term memory (LSTM) network as the greatest promise, ISM061-018-2 and ISM061-022, were character-
the classical model; and (3) Chemistry42 for validation (Fig. 1). The ized (structures in Fig. 2a,e, respectively). Their assessment involved
quantum component, depicted in Extended Data Fig. 3c, was a QCBM a two-stage process; the SPR determined their binding affinities and
that generated samples from quantum hardware in every training cell-based assays were subsequently used to gauge biological efficacy.
epoch and was trained with a reward value, P(x) = softmax(R(x)) cal- The compound ISM061-018-2, engineered through our hybrid quantum
culated using Chemistry42 or a local filter. We note that, in practice, model (Fig. 2b and Supplementary Table 6), demonstrated substantial
the reward function can be tailored to a user’s specific criteria, includ- binding affinity to KRAS-G12D, registered at 1.4 μM. To delve deeper
ing accounting for off-target effects by evaluating multiple docking into this molecule’s effectiveness across a spectrum of KRAS mutants,
scenarios. This recurrent sampling, training and validation process we commenced an extensive series of tests using a cell-based assay.
formed a cycle that continuously improved the generated molecular Specifically, we evaluated the molecule’s performance in a biological
structures targeting the KRAS protein. The final step was experi- context using a commercial cell viability assay (CellTiter-Glo, Promega)
mental validation. From our trained models, we sampled 1 million in conjunction with MaMTH-DS (mammalian membrane two-hybrid
compounds using three different models (selection workflow in drug screening), an advanced split-ubiquitin-based platform for the
Fig. 1). We used Chemistry42 to screen these samples for pharmaco- real-time detection of small molecules targeting specific cellular
logical viability and ranked them on the basis of their docking scores interactions28–35.
(protein–ligand interaction (PLI) score). The top 15 candidates were The biological activity of ISM061-018-2 was rigorously tested.
synthesized and tested using surface plasmon resonance (SPR) and Importantly, it demonstrated no detrimental impact on the viability of
cell-based assays. This study highlights the promising intersection HEK293 cells, even when expressing KRAS wild type (WT) or KRAS-G12V
of quantum computing and drug discovery, marking a quantum bait in MaMTH-DS format and being subjected to concentrations
computer’s first experimental hit. as high as 30 μM for 18–20 h, showing that the compound did not
Before initiating our campaign to design additional KRAS inhibi- possess any general, nonspecific toxicity (Fig. 2d). Subsequent testing
tors, we aimed to compare our hybrid quantum–classical approach to using MaMTH-DS across a spectrum of cell lines expressing various
established classical algorithms. We benchmarked our model, targeting KRAS baits (WT and five clinically important oncogenic mutants)
three different proteins, against the state of the art using the Tartarus in combination with Raf1 ‘prey‘ (a recognized KRAS effector) revealed
benchmarking suite for drug discovery26. The results indicate that, a dose-responsive inhibition of interactions, with half-maximal
while few classical models achieved high success rates and docking inhibitory concentration (IC50) values in the micromolar range
scores, QCBM–LSTM excelled by generating numerous high-quality (Fig. 2c and Supplementary Table 3). The compound’s activity was
samples. It achieved a high success rate and a docking score comparable not specific to mutants, as it targeted both WT and mutant inter-
to the best, as demonstrated in Supplementary Table 1. As illustrated actions with similar efficacy. It also showed comparable effective-
in Extended Data Fig. 1a and Supplementary Table 2, we evaluated the ness in disrupting the interactions of WT NRAS and HRAS baits with
impact of incorporating a quantum prior (QCBM/multibase (M)QCBM– Raf1 prey (Fig. 2c). However, it had no effect on the interaction of a
LSTM) versus a fully classical model (vanilla LSTM). We observed that completely unrelated artificial bait–prey control pair (consisting
a quantum prior enhanced the success rate, as measured by the pro- of membrane-anchored ALFA tag bait and nanobody ALFA prey)36,
portion of molecules meeting the criteria set by the same filters. Our supporting the biological specificity of the interaction (Fig. 2c).
observations revealed that the use of QCBM–LSTM (Supplementary Collectively, these results support a potential pan-Ras activity of
Table 1) offered a 21.5% improvement in passing filters that assessed ISM061-018-2.
the synthesizability and stability of the generated molecules, indicat- ISM061-022, as illustrated in Fig. 2e, also stood out as a compound
ing a higher quality of generated structures. Additionally, we analyzed of promise, particularly because of its selectivity toward certain KRAS
how the size of the prior (number of qubits) affected sample quality. mutants. Within our in vitro examination, ISM061-022 demonstrated
We found that more qubits improved the success rates for molecule a concentration-dependent inhibition of KRAS interactions with IC50
generation, suggesting that larger quantum models could enhance values in the micromolar range (Fig. 2f and Supplementary Table 4),
molecular design. Our findings suggest that the success rate correlates while manifesting only a mild, general impact on cell viability at higher
approximately linearly with the number of qubits, as illustrated in concentrations over an 18–20-h exposure (Fig. 2g). The observed inhibi-
Extended Data Fig. 1c. tion mirrored that of ISM061-018-2 yet displayed enhanced selectivity
We believe that the improvement in our hybrid classical–quantum toward certain KRAS mutants, particularly KRAS-G12R and KRAS-Q61H,
approach stemmed from the quantum effects, such as superposi- which were most receptive to the compound’s action (Fig. 2f). Diverging
tion and entanglement, which allow QCBMs to explore and represent from ISM061-018-2, ISM061-022 did not show binding to KRAS-G12D

Nature Biotechnology
 Brief Communication https://doi.org/10.1038/s41587-024-02526-3

ISM061-018-2

a b SPR: ISM061-018-2 c KRAS WT KRAS G12D KRAS G12C

150 150 150

% Activity

% Activity

% Activity
100 100 100
3

2 50 50 50

RU
1
0 0 0
0 –2 0 2 –2 0 2 –2 0 2
log [uM] log [uM] log [uM]
–1
–50 0 50 100 150 200 250 300
KRAS G12V KRAS G12R KRAS Q61H
Times(s)
150 150 150

% Activity

% Activity

% Activity
100 100 100

KRAS WT KRAS G12V

Cell viability (% of control)
Cell viability (% of control)

150 150 50 50 50

0 0 0
100 100
–2 0 2 –2 0 2 –2 0 2
log [uM] log [uM] log [uM]
50 50

HRAS WT NRAS WT Artificial bait

0 0 150 150
–1 0 1 2 –1 0 1 2 150

log [uM] log [uM]

% Activity

% Activity

% Activity
100 100 100

50 50 50

0 0 0
–2 0 2 –2 0 2 –2 0 2
log [uM] log [uM] log [uM]

ISM061-022

e f KRAS WT KRAS G12D KRAS G12C

150
150 150
% Activity

% Activity

% Activity
100
100 100

50
50 50

0 0 0
–2 0 2 –2 0 2 –2 0 2
log [uM] log [uM] log [uM]

KRAS G12V KRAS G12R KRAS Q61H

150 150 150
% Activity

% Activity

% Activity

100 100 100

g KRAS WT KRAS G12V

50 50 50
Cell viability (% of control)

Cell viability (% of control)

150 150

0 0 0
100 100 –2 0 2 –2 0 2 –2 0 2
log [uM] log [uM] log [uM]
50 50
HRAS WT NRAS WT Artificial bait
200 200
0 0 150
–1 0 1 2 –1 0 1 2
150 150
% Activity

% Activity

% Activity

log [uM] log [uM] 100

100 100

50
50 50

0 0 0
–2 0 2 –2 0 2 –2 0 2
log [uM] log [uM] log [uM]

Fig. 2 | Pharmacological characterization of compounds ISM061-018-2 and s.d. Each graph is representative of n = 3 biological replicates, performed
ISM061-022 through SPR and cellular activity assays. a,e, Chemical structures under the same conditions. Analyses were carried out using GraphPad Prism.
of ISM061-018-2 (a) and ISM061-022 (e). b, SPR sensorgrams illustrating binding d,g, Results from CellTiter-Glo viability assays measuring the impact of ISM061-
effects of ISM061-018-2 on KRAS protein. c,f, MaMTH-DS dose–response curves 018-2 (d) and ISM061-022 (g) on cellular proliferation over a concentration
illustrate the binding kinetics and effects of ISM061-018-2 (c) and ISM061-022 range from 123 nM to 30 μM, demonstrating that they do not display general,
(f) on various KRAS proteins, as well as NRAS, HRAS and artificial bait, with nonspecific toxicity. Data points are presented as averages from n = 3 technical
activities measured across concentrations from 4 nM to 30 μM. Data points are replicates with error bars showing the s.d. Each graph is representative of n = 3
presented as averages from n = 4 technical replicates with error bars showing the biological replicates performed under the same conditions.

Nature Biotechnology
Brief Communication https://doi.org/10.1038/s41587-024-02526-3

according to SPR. The compound also demonstrated activity against This synergistic approach could potentially lead to the discovery of
WT HRAS and NRAS, although it was less potent against HRAS. Notably, therapeutics that might be overlooked by other methods, while also
the compound had an unusual effect on our artificial control inter- reducing the preclinical drug discovery phase from several years to
action pair, leading to a distinct dip to approximately 50% residual just a few months.
activity at concentrations of ~250–500 nM, before climbing again and
ultimately leading to mild enhancement of the interaction at higher Online content
micromolar concentrations (Fig. 2f). This distinct pattern suggests Any methods, additional references, Nature Portfolio reporting sum-
an alternative mode of action for ISM061-022, revealing at least some maries, source data, extended data, supplementary information,
degree of nonspecific activity, although partial specificity for mutant acknowledgements, peer review information; details of author contri-
KRAS protein does still appear to be in evidence. Further investiga- butions and competing interests; and statements of data and code avail-
tion into the mechanism of action of this molecule will be the focus ability are available at https://doi.org/10.1038/s41587-024-02526-3.
of future study.
Overall, these live-cell experimental observations underscore the References
robustness of our approach, effectively identifying small-molecule 1. Petrova, E. Innovation in the pharmaceutical industry: the process
candidates with biological activity. While these molecules are certainly of drug discovery and development. In Innovation and Marketing
at an early stage and in need of further refinement, our results under- in the Pharmaceutical Industry: Emerging Practices, Research, and
line the potential of our methodology to address and surmount the Policies (eds Ding, M. et al.) 19–82 (Springer, 2014).
complexities inherent in targeting clinically challenging biomolecules. 2. DiMasi, J. A., Grabowski, H. G. & Hansen, R. W. Innovation in the
To further characterize ISM061-018-2’s interaction with KRAS-G12D, pharmaceutical industry: new estimates of R&D costs. J. Health
we used protein-detected nuclear magnetic resonance (NMR) spec- Econ. 47, 20–33 (2016).
troscopy. 15N-labeled KRAS-G12D (residues 1–169) protein was used to 3. Sun, D., Gao, W., Hu, H. & Zhou, S. Why 90% of clinical drug
record 1H–15N transverse relaxation optimized spectroscopy (TROSY) development fails and how to improve it? Acta Pharm. Sin. B 12,
heteronuclear single quantum coherence (HSQC) spectra in the pres- 3049–3062 (2022).
ence and absence of ISM061-018-2 under guanosine diphosphate 4. Zhavoronkov, A. et al. Deep learning enables rapid identification
(GDP)-loaded conditions (Extended Data Fig. 2a). Substantial chemical of potent ddr1 kinase inhibitors. Nat. Biotechnol. 37, 1038–1040
shift perturbations (CSPs) were observed, indicating fast exchange (2019).
between free and bound states, with some residues showing intensity 5. Stokes, J. et al. A deep learning approach to antibiotic discovery.
loss because of exchange broadening. Mapping the weighted CSP Cell 180, 688–702 (2020).
values onto the Protein Data Bank (PDB) structure of GDP-bound 6. Ren, F. et al. AlphaFold accelerates artificial intelligence powered
KRAS-G12D (Extended Data Fig. 2b) revealed notable perturbations drug discovery: efficient discovery of a novel CDK20 small
in the P-loop, α2 helix, switch II region and nearby β4 sheet residues, molecule inhibitor. Chem. Sci. 14, 1443–1452 (2023).
suggesting a binding mode within the switch II pocket. Additional CSPs 7. Nigam, A. K., Pollice, R., Friederich, P. & Aspuru-Guzik, A. Artificial
in the α1 helix, switch I and α4 regions suggest conformational changes design of organic emitters via a genetic algorithm enhanced by
upon ligand binding. Although direct interaction with D12 (ref. 21), a a deep neural network. Chem. Sci. 15, 2618–2639 (2024).
residue critical for selectivity in inhibitors such as MRTX1133, was not 8. Sánchez-Lengeling, B. & Aspuru Guzik, A. Inverse molecular
confirmed, CSPs in the P-loop residues (G10 and A11) imply possible spe- design using machine learning: generative models for matter
cific interactions. These NMR results strongly support ISM061-018-2’s engineering. Science 361, 360–365 (2018).
binding to the switch II pocket of KRAS-G12D, further corroborat- 9. Pollice, R. et al. Data-driven strategies for accelerated materials
ing its proposed mechanism of action. However, cocrystallization design. Acc. Chem. Res. 54, 849–860 (2021).
studies are necessary to fully elucidate the binding residues. In sum- 10. Nigam, A. K. et al. Assigning confidence to molecular property
mary, current quantum hardware is proving crucial for practical drug prediction. Exp. Opin. Drug Discov. 16, 935–949 (2021).
discovery applications, preempting the need for fully fault-tolerant 11. Bohacek, R. S., McMartin, C. & Guida, W. C. The art and practice of
quantum systems in the future. Quantum algorithms have the potential structure-based drug design: a molecular modeling perspective.
to design small molecules that better align with the distribution of the Med. Res. Rev. 16, 3–50 (1996).
provided training set, achieving outcomes that are challenging for 12. Santagati, R. et al. Drug design on quantum computers. Nat. Phys.
classical methods because of their ability to more effectively explore 20, 549–557 (2024).
and represent complex, high-dimensional probability distributions. 13. Li, J., Topaloglu, R. O. & Ghosh, S. Quantum generative models
Using just 16 qubits, our algorithm demonstrates how quantum com- for small molecule drug discovery. IEEE Trans. Quantum Eng. 2,
puting can enhance classical computation, leading to more efficient 1–8 (2021).
algorithms. Our hybrid quantum–classical model surpasses classical 14. Batra, K. et al. Quantum machine learning algorithms for drug
models in performance and matches up well against the state of the discovery applications. J. Chem. Inf. Model. 61, 2641–2647
art according to the Tartarus benchmark. We applied this model to (2021).
design inhibitors targeting the complex KRAS protein, synthesizing 15. Benedetti, M., Lloyd, E., Sack, S. & Fiorentini, M. Parameterized
and evaluating 15 promising ligands. Among these, ISM061-018-2 and quantum circuits as machine learning models. Quantum Sci.
ISM061-022 stood out, showing inhibitory effects on KRAS variants Technol. 4, 043001 (2019).
with ISM061-018-2 displaying a high binding affinity of 1.4 μM and 16. Hibat-Allah, M., Mauri, M., Carrasquilla, J. & Perdomo-Ortiz, A.
broad activity across KRAS mutants. These results, derived from our A framework for demonstrating practical quantum advantage:
hybrid approach, indicate superior efficacy to fully classical models racing quantum against classical generative models. Commun.
and introduce potential chemotypes that expand the chemical space Phys. 7, 68 (2024).
explored in KRAS inhibitor design. While these findings are encour- 17. McClean, J. R., Boixo, S., Smelyanskiy, V., Babbush, R. & Neven, H.
aging, they stop short of definitively proving a ‘quantum advantage’, Barren plateaus in quantum neural network training landscapes.
achieving results unattainable by classical methods within a reasonable Nat. Commun. 9, 4812 (2018).
time frame. Future work will focus on exploring transformer-based 18. Gili, K., Hibat-Allah, M., Mauri, M., Ballance, C. J. & Perdomo-Ortiz, A.
generative models, coupled with increasing the number of qubits, Do quantum circuit born machines generalize? Quantum Sci.
to further enhance the quality and diversity of generated molecules. Technol. 8, 015021 (2023).

Nature Biotechnology
Brief Communication https://doi.org/10.1038/s41587-024-02526-3

19. Rudolph, M. S. et al. Generation of high-resolution handwritten 32. Scheper, W., Thaminy, S., Kais, S., Stagljar, I. & Römisch, K.
digits with an ion-trap quantum computer. Phys. Rev. X 12, 031010 Coordination of N-glycosylation and protein translocation across
(2022). the endoplasmic reticulum membrane by Sss1 protein.
20. Zeng, Q., Ge, H.-Y., Chen, G. & Zhou, N. Conditional quantum J. Biol. Chem. 278, 37998–38003 (2003).
circuit born machine based on a hybrid quantum-classical 33. Benleulmi-Chaachoua, A. et al. Protein interactome mining
framework. Physica A Stat. Mech. Appl. 608, 128693 (2023). defines melatonin MT1 receptors as integral component of
21. Wang, X. et al. Identification of MRTX1133, a noncovalent, potent, presynaptic protein complexes of neurons. J. Pineal Res. 60,
and selective KRasG12D inhibitor. J. Med. Chem. 65, 3123–3133 (2022). 95–108 (2016).
22. Mao, Z. et al. KRas(G12D) can be targeted by potent inhibitors via 34. Lopes, J. P. et al. The role of Parkinson’s disease-associated
formation of salt bridge. Cell Discov. 8, 5 (2022). receptor GPR37 in the hippocampus: functional interplay
23. Gorgulla, C. et al. Virtualflow 2.0—the next generation drug with the adenosinergic system. J. Neurochem. 134, 135–146
discovery platform enabling adaptive screens of 69 billion (2015).
molecules. Preprint at bioRxiv https://doi.org/10.1101/ 35. Grozavu, I. et al. D154Q mutation does not alter KRas dimerization.
2023.04.25.537981 (2023). J. Mol. Biol. 434, 167543 (2022).
24. Nigam, A. K., Pollice, R., Krenn, M., Gomes, G. & Guzik, A. A. 36. Götzke, H. et al. The ALFA-tag is a highly versatile tool for
Beyond generative models: superfast traversal, optimization, nanobody-based bioscience applications. Nat. Commun. 10,
novelty, exploration and discovery (STONED) algorithm for 4403 (2019).
molecules using SELFIES. Chem. Sci. 12, 7079–7090 (2021).
25. Ivanenkov, Y. A. et al. Chemistry42: an AI-driven platform for Publisher’s note Springer Nature remains neutral with regard to
molecular design and optimization. J. Chem. Inf. Model. 63, jurisdictional claims in published maps and institutional affiliations.
695–701 (2023).
26. Nigam, A. et al. Tartarus: A benchmarking platform for realistic Open Access This article is licensed under a Creative Commons
and practical inverse molecular design. In Advances in Neural Attribution-NonCommercial-NoDerivatives 4.0 International License,
Information Processing Systems 36, 3263–3306 (2023). which permits any non-commercial use, sharing, distribution and
27. Schuld, M., Sinayskiy, I. & Petruccione, F. An introduction to reproduction in any medium or format, as long as you give appropriate
quantum machine learning. Quantum Mach. Intell. 3, 28 (2021). credit to the original author(s) and the source, provide a link to the
28. Saraon, P. et al. A drug discovery platform to identify compounds Creative Commons licence, and indicate if you modified the licensed
that inhibit EGFR triple mutants. Nat. Chem. Biol. 16, 577–586 (2020). material. You do not have permission under this licence to share
29. Stagljar, I., Korostensky, C., Johnsson, N. & te Heesen, S. A genetic adapted material derived from this article or parts of it. The images
system based on split-ubiquitin for the analysis of interactions or other third party material in this article are included in the article’s
between membrane proteins in vivo. Proc. Natl Acad. Sci. USA 95, Creative Commons licence, unless indicated otherwise in a credit
5187–5192 (1998). line to the material. If material is not included in the article’s Creative
30. Petschnigg, J. et al. The mammalian-membrane two-hybrid assay Commons licence and your intended use is not permitted by statutory
(MAMTH) for probing membrane–protein interactions in human regulation or exceeds the permitted use, you will need to obtain
cells. Nat. Methods 11, 585–592 (2014). permission directly from the copyright holder. To view a copy of this
31. Saraon, P. et al. Chemical genetics screen identifies COPB2 licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
tool compounds that alters ER stress response and induces RTK
dysregulation in lung cancer cells. J. Mol. Biol. 433, 167294 (2021). © The Author(s) 2025

Nature Biotechnology
Brief Communication https://doi.org/10.1038/s41587-024-02526-3

Methods at capturing the distribution of the provided molecule set would exhibit
Benchmark setup a correspondingly high success rate in generating novel molecules
Our benchmark used both classical and quantum hardware. Our classi- without structural violations. Our observations indicated that only a
cal computational setup was based on a cluster equipped with graphics few generative models demonstrated a high success rate. However, the
processing unit (GPU) nodes. This cluster consisted of two GPU nodes, QCBM–LSTM model was very strong in producing a substantial num-
each with specific features. These features included two AMD EPYC ber of high-quality samples that successfully meet the filter criteria,
7V13 64-core processors, resulting in a total of 128 central processing as evidenced by the elevated success rate depicted in Supplementary
unit cores per node. In addition, each node was equipped with 512 GB Table 1. Consequently, we believe that the incorporation of a quantum
of random-access memory (RAM). The nodes also contain eight AMD prior leads to improved distribution matching. We further benchmark
Instinct MI100 GPUs, each with a GPU RAM of 32 GB. For the classical the influence of a classical–quantum prior in the subsequent sec-
training, we used four of these GPUs in parallel (that is, one GPU node). tion. Moreover, our analysis revealed that, for the PDB 4LDE target,
Furthermore, we used an Nvidia GPU (RTX3090Ti) to facilitate our our model generated the highest-scoring molecules relative to other
classical–quantum simulations. For the quantum hardware setup, we generative models. While the docking scores for the remaining two
used the Guadalupe quantum system, equipped with 16 qubits and a targets were not as high as those produced by classical algorithms, we
Falcon r4P processor type. Our QCBM model, accompanied by an error speculate that incorporating a docking-score-based reward, in conjunc-
correction circuit, was executed on this quantum processor. tion with the filter success rate, could potentially improve our results.
Regarding software, we used several packages provided by Zapata
AI under the Qml core agreements. We implemented our variational Benchmarking of prior distributions. To evaluate the impact of prior
quantum circuit and classical LSTM model using the Qml Core Python selection on the quality of the molecules generated by our model, we
package. We used the STONED–SELFIES and VirtualFlow 2.0 packages trained four distinct model variants, each incorporating different
to prepare a diverse dataset. Additionally, we used RDkit and Insilico priors (Extended Data Fig. 1b). Specifically, we examined a QCBM
APIs to compute the reward value and conduct some postprocessing prior and implemented it on both a quantum simulator and a hard-
analyses. The QCBM model underwent a training regimen spanning ware backend, in contrast with an MQCBM operating exclusively on
30 epochs. In contrast, the LSTM model was trained over a total of a quantum simulator and a classical LSTM model devoid of quantum
40 epochs. priors. These models were tasked with designing KRAS inhibitors,
We used the Optuna platform to optimize the hyperparameters using a meticulously curated dataset of over 1 million molecules (Fig. 1).
in the benchmarking. We ran Optuna tuning for 100 trials for each Extended Data Figure 1b showcases the optimal results obtained follow-
model to determine the optimal number of QCBM layers, number of ing a comprehensive optimization of the corresponding architectures
LSTM layers and embedding dimensions. Additionally, we tuned the using Optuna52. We assessed the quality of the generated molecules
sampling temperature, which defines the balance between determin- using two distinct sets of criteria: one derived from Tartarus26, termed
ism and stochasticity in the model, particularly between the prior input the ‘local filter’, and a more stringent set provided by Chemistry42,
and the LSTM output. termed the ‘Chemistry42 filter’. In both assessments, we observed that
incorporating a quantum prior enhanced the success rate, as gauged
Computational benchmarks: classical versus quantum models by the proportion of molecules satisfying the criteria set by the two
Tartarus benchmark. We used the Tartarus platform26 to benchmark filters. Furthermore, using the top model from each prior category, we
our proposed QCBM–LSTM methodology against an array of classical sampled 5,000 molecules that successfully met the filter criteria and
state-of-the-art models, including REINVENT37, SMILES–VAE38, SELFIES– examined their respective docking scores (Supplementary Table 2).
VAE39, MoFlow40, SMILES–LSTM–HC41,42, SELFIES–LSTM–HC, GB–GA43 Intriguingly, these molecules displayed comparably high docking
and JANUS44. The study focused on three protein targets selected from scores as determined by QuickVina 2 and the PLI score, as evaluated by
the Tartarus dataset: (1) PDB 1SYH, an ionotropic glutamate receptor Chemistry42. Additionally, the synthesized molecules demonstrated
associated with neurological and psychiatric disorders such as consistent metrics across various parameters, including the diversity
Alzheimer disease, Parkinson disease and epilepsy45; (2) PDB 6Y2F, the fraction, uniqueness fraction, Chemistry42 reward and Chemistry42
main protease of severe acute respiratory syndrome coronavirus 2, synthetic accessibility score25.
crucial for its RNA translation46; and (3) PDB 4LDE, the β2-adrenoceptor Encouraged by our observation that quantum priors enhance
G-protein-coupled receptor, a cell-membrane-spanning receptor that molecule quality, we further investigated the influence of the number
binds to adrenaline, a hormone implicated in muscle relaxation and of qubits used in modeling priors on the quality of generated molecules
bronchodilation47. For each target, we had a dual objective: to generate (Extended Data Fig. 1d). Specifically, we analyzed the percentage of
novel molecules that exhibit strong binding affinity to the specified 5,000 uniquely generated random molecules that satisfied a series of
proteins, as determined by active sites assigned by Tartarus, and to local filters. Interestingly, our findings revealed that the success rate
minimize the docking score using QuickVina 2 (ref. 48). Additionally, correlated roughly linearly with the number of qubits used in modeling
these molecules were required to pass a comprehensive set of filters the prior, indicating a direct relationship between the complexity of
designed to eliminate reactive, unsynthesizable or unstable groups, the quantum model and the effectiveness in generating high-quality
thereby streamlining the drug discovery process. The top-performing molecules. This trend underscores the potential of increasing qubit
molecules, after filtering, were subjected to a refined rescoring using numbers in quantum models to improve molecular design outcomes
a more precise scoring function provided by SMINA49, at an increased systematically.
level of exhaustiveness.
We conducted experiments using the QCBM with 16 qubits as SPR conditions
a quantum prior and the LSTM as a classical model. The local filter A Biacore 8K system was used for all experiments. For preliminary
from the Tartarus paper served as the reward function to train the compound screening, N-terminal biotinylated KRAS-G12D protein
QCBM. As recommended by Tartarus, our models were trained on a (synthesized by VIVA Biotech; purity ≥ 95%) was captured on a sen-
subset of 150,000 molecules from the Developmental Therapeutics sor chip SA (GE Healthcare) at a density of about 2,000 RU. Protein
Program open compound collection50,51, referred to as DATASET in immobilization was conducted using 1× HBS-EP+, 2 mM TCEP and 2%
Supplementary Table 1. Notably, all 150,000 structures underwent a DMSO as a running buffer. Protein was injected for 70 s at a flow rate
rigorous screening process using structural filters to eliminate reactive, of 5 μl min−1. The protein concentration was 5 μg ml−1. We performed
unsynthesizable or unstable groups. As such, generative models adept an initial screening of compounds prepared samples by serial twofold

Nature Biotechnology
Brief Communication https://doi.org/10.1038/s41587-024-02526-3

dilutions from 200 μM to 0.39 μM in 1× HBS-EP+, 2 mM TCEP and 2% sodium phosphate pH 8.0, 500 mM NaCl, 10 mM imidazole, 1 mM
DMSO. Samples were injected for 60 s at a flow rate of 30 μl min−1 and 2-mercaptoethanol and 5% (v/v) glycerol) containing PMSF and benza-
dissociation time of 180 s. A Biacore 8K machine was used to carry out midine. Protein was purified over a His trap column (Cytiva) following
the SPR experiments and subsequent data analysis. standard Ni-affinity protocols (wash with 10 mM and 20 mM imidazole
and then elute with 300 mM imidazole) and the His-tag was removed
MaMTH-DS dose–response assays by TEV cleavage. KRAS-G12D was further purified by size-exclusion
MaMTH-DS FLP HEK293 reporter cell lines28 stably expressing KRAS chromatography using a Superdex 75 Increase column (10/300 GL)
(WT or mutant), HRAS, NRAS or artificial membrane-anchored ALFA with PBS buffer (137 mM NaCl, 2.7 mM KCl, 10 mM Na2HPO4 and 2 mM
tag bait alongside Raf1 (for Ras baits) or nanobody ALFA (for artificial KH2PO4, pH 7.4) containing 10 mM EDTA. The samples were prepared
bait) preys were seeded into 384-well white-walled, flat-bottomed, by mixing 30 μM of protein with 200 μM ISM018-2 in the presence of
tissue-culture-treated microplates (Greiner, 781098) at a concentra- 2% DMSO-d6, 5 mM GDP and 15 mM MgCl2 in the buffer. The purifica-
tion of 100,000 cells per ml (50 μl total volume per well) in DMEM, 10% tion protocols were adapted from previously published protocols for
FBS and 1% penicillin–streptomycin. Seeding was performed using a KRAS-G12C (ref. 53) and KRAS-G12D (ref. 54).
MultiFlo-FX multimode liquid dispenser (BioTek). Plates were left at
room temperature for 30–60 min following seeding before transfer NMR experiments and analysis
1
to a Heracell 150i incubator (Thermo Fisher Scientific) and growth at H–15N TROSY HSQC experiments were recorded at 298 K on an
37 °C in 5% CO2 for 3 h. After growth, 10 μl of DMEM, 10% FBS and 1% 800-MHz Bruker spectrometer equipped with an AVANCE III console
penicillin–streptomycin supplemented with 3 μg ml−1 tetracycline and a cryogenically cooled probe. The acquired spectra were processed
(to induce bait and prey expression; BioShop, TET701) and 60 ng ml−1 using NMRPipe55 and were analyzed using ccpnmr 2.0 (ref. 56). The
epidermal growth factor (to stimulate Ras signaling; Sigma, E9644) was chemical shifts were transferred from Biological Magnetic Resonance
added to each well using a multichannel pipette. As appropriate, a 6× Bank 27719 (ref. 54). The CSPs on both 1H and 15N dimensions were used
concentration of drug (or DMSO only) was also included in the medium, to calculate the weighted CSP57 and these values were then plotted
with all lower concentrations produced by serial dilution starting from onto the PDB structure using PyMol to map the regions that took part
the highest concentration solution. Plates were then grown overnight in ligand binding or underwent conformational changes upon ligand
(18–20 h) at 37 °C in 5% CO2. A luciferase assay was performed the next binding.
day using 10 μl of 20 μM native coelenterazine substrate (Nanolight,
303) per well. Luminescence was measured using a Clariostar plate Experimental validation methods
reader (BMG Labtech) with a gain of 3,200–3,800 and a 1-s integration Chemistry42 methods: after screening and selection of promising
time. All data analysis was performed using Microsoft Excel and Graph- candidate structures for synthesis. Our selection process relied on a
Pad Prism. Curve fits were performed in Prism (nonlinear regression) tiered system of the following criteria:
using log(inhibitor) versus response curves, with a variable slope (four
• S tructural and compositional parameter filters, such as hydro-
parameters) bottom-constrained to zero.
gen bond donor count and aromatic atom fraction.
• Property evaluation for molecular weight, lipophilicity and
Cell viability assay
other physicochemical traits.
MaMTH-DS FLP HEK293 reporter cell lines28 stably expressing KRAS
• Medicinal chemistry filters to exclude problematic structural
(WT or G12V mutant) bait alongside Raf1 prey were seeded into 96-well
motifs.
white-walled, μCLEAR flat-bottomed, tissue-culture-treated plates
• A synthetic accessibility assessment based on the ReRSA
(Greiner, 655098) at 40,000 cells per well in DMEM, 10% FBS and 1%
model58.
penicillin–streptomycin (60 μl total volume per well). Seeding was
• Three-dimensional pharmacophoric analysis in reference to an
performed using a MultiFlo-FX multimode liquid dispenser (BioTek).
X-ray cocrystal structure (PDB 7EW9) (Extended Data Fig. 4).
Plates were left at room temperature for 30–60 min following seeding
• PLI scoring to estimate binding efficiency.
before transfer to a Heracell 150i incubator (Thermo Fisher Scientific)
• An overall reward calculation integrating the above scores
and growth at 37 °C in 5% CO2 for 3 h. After growth, 30 μl of a 3× con-
centration of the drug (or DMSO only) in DMEM, 10% FBS and 1% peni- For synthesis candidacy, we applied more demanding conditions,
cillin–streptomycin was added to wells, with all lower concentrations ensuring compounds met the following enhanced benchmarks:
produced by serial dilution starting from the highest concentration • Clearance of all Chemistry42 filters.
solution (final drug concentration: 30 μM to 123 nM). Plates were then • An aggregate reward value exceeding 0.7.
grown overnight (18–20 h) at 37 °C in 5% CO2. The effect of the drug on • A PLI score indicative of strong KRAS binding (<−8 kcal mol−1).
cell viability was assessed by the CellTiter-Glo luminescent cell viability • A pharmacophore match score above 0.7.
assay from Promega (G7570). Briefly, 90 μl of the CellTiter-Glo reagent • Favorable synthetic accessibility with a ReRSA score under 5.
was added directly into each well following 30-min equilibration of the
plate at room temperature. Contents of the wells were mixed on an After screening, molecules were clustered and ranked within clusters
orbital shaker for 2 min and plates were then incubated at room tem- by chemical similarity, allowing for expert analysis to further prioritize
perature for 10 min to stabilize the luminescence signal. Luminescence on the basis of novelty and structural intricacy, culminating in a selec-
was measured using a Clariostar plate reader (BMG Labtech) with a gain tion of 100–150 molecules for potential synthesis and subsequent
of 3,600 and a 1-s integration time. Values represent the mean ± s.d. of examination.
three replicates for each tested drug concentration. All data analysis
was performed using Microsoft Excel and GraphPad Prism. Experimental evaluation methods of generated compounds. From
the pool of identified structures, we synthesized and characterized 15
Protein purification and NMR: sample preparation compounds. Detailed methodologies of this process are elaborated in
A construct encoding N-His–TEV (tobacco etch virus)–KRAS-G12D the Supplementary Information. The molecular structures of the two
was transformed into BL21(DE3) cells. The cells were grown in minimal most promising compounds (ISM061-018-2 and ISM061-022) are show-
medium containing 1 g l−1 [15N]H4Cl and 4 g l−1 glucose to an optical cased in Fig. 2a,e. Each synthesized compound underwent a rigorous
density at 600 nm of ~0.7–1.0 and induced with 1 mM IPTG for 16 h two-phase evaluation; their binding affinities were determined using
at 16 °C. Cells were pelleted and resuspended in lysis buffer (20 mM SPR and their biological efficacies were gauged through cell-based

Nature Biotechnology
Brief Communication https://doi.org/10.1038/s41587-024-02526-3

assays. Notably, compound ISM061-018-2, engineered through our Given the dataset’s limited size, we opted to expand it to improve the
hybrid quantum model (Supplementary Table 3), demonstrated a robustness of our model during training.
substantial binding affinity to KRAS-G12D, registered at 1.4 μM. To
delve deeper into this molecule’s effectiveness across a spectrum STONED–SELFIES. We used the STONED–SELFIES24 algorithm (https://
of KRAS mutants, we commenced an extensive series of tests using github.com/aspuru-guzik-group/stoned-selfies) to mine our initial
a cell-based assay. Specifically, we evaluated the molecule’s perfor- set of 650 molecules. For a given molecule in SMILES format, we first
mance in a biological context using a commercial cell viability assay randomized the string using RDKit. These randomized strings were
(CellTiter-Glo, Promega) in conjunction with MaMTH-DS, an advanced then converted into SELFIES. Each SELFIES string underwent mutations
split-ubiquitin-based platform for the real-time detection of small (in the form of character deletions, replacements and additions) up
molecules targeting specific cellular interactions28–34. to 500 times. Subsequently, the synthesizability and stability of the
Furthermore, the biological activity of ISM061-018-2 was rigor- mutated strings were assessed using Chemistry42. We generated 850,000
ously tested. Importantly, it demonstrated no detrimental impact on molecules, which served as the training set for our generative models.
the viability of HEK293 cells, even when expressing either KRAS WT
or KRAS-G12V bait in MaMTH-DS format and being subjected to con- Virtual screening process. VirtualFlow 2.0 (ref. 23) was used to iden-
centrations as high as 30 μM for 18–20 h, showing that the compound tify additional molecules predicted to bind to KRAS-G12D. The adaptive
did not possess any general, nonspecific toxicity (Fig. 2d). Subsequent target-guided (ATG) method performed the virtual screening in two
testing using MaMTH-DS across a spectrum of cell lines expressing vari- stages. In the first stage, the ATG prescreen was performed, in which
ous KRAS baits (WT and five clinically important oncogenic mutants) a spare version of the 69 billion REAL space from Enamine (version
in combination with Raf1 prey (a recognized KRAS effector) revealed a 2022q12) was screened. In the second stage, the most potent tranches
dose-responsive inhibition of interactions, with IC50 values in the micro- of ligands were screened in full, amounting to 100 million ligands. The
molar range (Supplementary Table 3). The compound’s activity was not docking program used was QuickVina 2 (ref. 48), with exhaustiveness
specific to mutants, as it targeted both WT and mutant interactions with set to 1 in both stages of the screen. The screen was carried out in the
similar efficacy. It also showed comparable effectiveness in disrupting Amazon Web Services cloud computing platform. The protein struc-
the interactions of WT NRAS and HRAS baits with Raf1 prey (Fig. 2c). ture used in the screen was PDB 5US4 (ref. 59), which was prepared
However, it had no effect on the interaction of a completely unrelated before the virtual screen with Schrödinger’s protein preparation wizard
artificial bait–prey control pair (consisting of membrane-anchored (addition of hydrogens and protonation state prediction). The size of
ALFA tag bait and nanobody ALFA prey)36, supporting the biological the docking box was 14 × 14 × 20 Å3.
specificity of the interaction (Fig. 2c). Collectively, these results sup-
port a potential pan-Ras activity of ISM061-018-2. Quantum-assisted algorithm
ISM061-022 (Fig. 2e) also stood out as a compound of promise, As shown in Fig. 1, our quantum-assisted model was a hybrid algorithm
particularly because of its selectivity toward certain KRAS mutants. composed of both quantum and classical generative components. The
Within our in vitro examination, ISM061-022 demonstrated a quantum generative model used a QCBM model while the classical
concentration-dependent inhibition of KRAS interactions (Fig. 2f), component used an LSTM model. Extended Data Fig. 3 illustrates the
while manifesting only a mild, general impact on cell viability at higher flowchart of our proposed generative model.
concentrations over an 18–20-h exposure (Fig. 2g). The observed inhibi- Within this model, we used Chemistry42 and a local filter to vali-
tion mirrored that of ISM061-018-2 yet displayed enhanced selectivity date sample generation at each step, which was then used to train
toward certain KRAS mutants, particularly KRAS-G12R and KRAS-Q61H, the QCBM model. The QCBM model, a quantum circuit model, was
which were most receptive to the compound’s action (Fig. 2f and Sup- executed on a quantum processing unit. Subsequently, samples from
plementary Table 4). Diverging from ISM061-018-2, ISM061-022 did the trained QCBM were fed into the LSTM model, which generated
not show binding to KRAS-G12D. The compound also demonstrated sequences on the basis of these samples. The reward value for each
activity against WT HRAS and NRAS, although it was less potent against sample was computed at every step using the local filter until epoch
HRAS. Notably, the compound had an unusual effect on our artificial 20, after which we selected Chemistry42. This reward value was then
control interaction pair, leading to a distinct dip to approximately 50% used to train our quantum generative model. During the first epoch,
residual activity at concentrations of 250–500 nM, before climbing no rewards were available; hence, the algorithm sampled from the
again and ultimately leading to mild enhancement of the interaction untrained QCBM model, designated as Xi. From the second epoch
at higher micromolar concentrations (Fig. 2f). This distinct pattern onward, rewards were computed, allowing us to calculate the softmax
suggests an alternative mode of action for ISM061-022, revealing at of the rewards for each Xi, where i ∈ [1, N]. The corresponding pseu-
least some degree of nonspecific activity, although partial specificity docode can be found in Algorithm 1.
for mutant KRAS protein does still appear to be in evidence. Further
investigation into the mechanism of action of this molecule will be the Quantum-computing-enhanced workflow
focus of future study. Our methodology encompassed a comprehensive workflow, extend-
In essence, these live-cell experimental observations underscore ing from data preparation to experimental validation, as delineated in
the robustness of our approach, effectively identifying small-molecule Fig. 1. This workflow was structured into three pivotal stages:
candidates with biological activity. This underlines the potential of our (1) Generation of training data. We initiated the process by con-
methodology to address and surmount the complexities inherent in structing a robust dataset for training our generative model to target
targeting clinically challenging biomolecules. the KRAS protein. The foundation of this dataset was approximately
This section explains the methods and workflow incorporated in 650 experimentally confirmed KRAS inhibitors, compiled through an
our proposed approach, offering a comprehensive understanding of extensive literature review60–63. Acknowledging the necessity of a more
the mechanisms used in our study. Figure 1 illustrates the workflow expansive dataset to develop a model for ligand design effectively, we
we used in our study. adopted a two-pronged approach: virtual screening and local chemical
space exploration. In the virtual screening phase, we used VirtualFlow
Data acquisition and preprocessing 2.0 (ref. 23) to screen 100 million molecules, applying Enamine’s REAL
Our preliminary dataset, sourced from Insilico Medicine, included library64 in conjunction with molecular docking techniques. The top
approximately 650 data points. These were selectively collated from 250,000 compounds from this screen, exhibiting the lowest docking
existing literature, specifically targeting the KRAS-G12D mutant (Fig. 1). scores, were subsequently integrated into our dataset. Complementing

Nature Biotechnology
Brief Communication https://doi.org/10.1038/s41587-024-02526-3

this, the local chemical space exploration was conducted using the initially configured with randomly assigned parameters |ψ(θ)⟩. These
STONED–SELFIES algorithm24, which was applied to the 650 experimen- parameters are subsequently calculated throughout the training pro-
tally derived hits. This algorithm distinctively introduces random point cess. The training of the QCBM model involves minimizing the exact
substitutions into the SELFIES representations39,65,66 of the molecules, negative log-likelihood (exact NLL) loss function.
thereby generating novel compounds that maintain a structural resem-
blance to the starting point. The resulting derivatives were filtered on Classical model: LSTM model. LSTM networks (Fig. 1) were used for
the basis of synthesizability, culminating in the addition of 850,000 the classical part of this architecture. LSTM is simple and has a good
molecules to our training set. record of learning the string pattern in natural language processing
(2) Generation of new molecules. Our approach was structured for a long time. LSTM networks are specialized recurrent neural
around the integration of three primary components: (1) the QCBM; networks capable of learning long-term dependencies in sequence
(2) the classical LSTM model; and (3) Chemistry42 for artificial- data67. They are particularly useful in applications where the context
intelligence-driven validation, as shown in Extended Data Fig. 3. The from earlier parts of the sequence is needed to interpret later parts,
QCBM generator16 used a 16-qubit IBM quantum processor with quan- such as in natural language processing and time-series forecasting68.
tum circuits to model complex data distributions. The integration The LSTM architecture consists of a chain of repeating modules
method of quantum priors into the LSTM architecture (Supplementary called cells. Each cell contains three gates that control the flow of
Fig. 4) involved merging molecular information encoded in SELFIES information:
and quantum data by addition or concatenation to form samples, X′ (t), Forget gate: This gate decides what information from the cell
1. 
which were then input into the LSTM cell. The quantum component state should be thrown away or kept. It takes the output of
(Supplementary Fig. 4) was a QCBM that generated samples from the previous LSTM cell and the current input and passes them
quantum hardware each training epoch and was trained with a reward through a sigmoid function, outputting a number between 0
value, P(x) = softmax(R(x)), calculated using Chemistry42 or a local and 1 for each number in the cell state, where 0 means ‘com-
filter. This cyclical sampling, training and validation process formed a pletely forget this’ and 1 means ‘completely keep this’.
loop aimed at continually improving the generated molecular struc- 2. Input gate: This gate updates the cell state with new informa-
tures for targeting KRAS. tion. It has two parts: a sigmoid layer called the input gate layer
(3) Experimental validation. The process of selecting experi- and a hyperbolic tangent layer. The sigmoid layer decides what
mental sample candidates is illustrated at the bottom of Fig. 1. After values to update and the hyperbolic tangent layer creates a vec-
training our model, we sampled 1 million compounds from each tor of new candidate values that could be added to the state.
prior model listed at the bottom of Fig. 1. These samples underwent 3. Output gate: This gate decides the next hidden state. The hid-
evaluation by Chemistry42, filtering out unsuitable compounds for den state contains information on previous inputs. The hidden
pharmacological purposes and ranking the remaining compounds by state is used to calculate the output of the LSTM and the next
their docking score (PLI score). Subsequently, 15 novel compounds hidden state.
were selected for synthesis and underwent SPR and cell-based assay
experiments. The following equations can describe the LSTM’s operations:

Forget gate: ft = σ(Wf ⋅ [ht−1 , xt ] + bf ) (1)

Algorithm 1. Algorithm 1 provides an overview of quantum-assisted
drug discovery using an LSTM framework. This pseudocode details the
iterative process, starting with the initialization of the LSTM and QCBM Input gate: it = σ(Wi ⋅ [ht−1 , xt ] + bi ) (2)
models, followed by the generation and validation of new compounds.
Valid compounds are subjected to a reward calculation and probability
Candidate values: Ct̃ = tanh(WC ⋅ [ht−1 , xt ] + bC ) (3)
assessment, which in turn inform the subsequent training of the QCBM
(Algorithm 2). This cycle continues until convergence, illustrating the
dynamic interplay between quantum predictions and LSTM-generated Update cell state: Ct = ft ⋅ Ct−1 + it ⋅ Ct̃ (4)
compounds underpinned by the Chemistry42 evaluation.
Output gate: ot = σ(Wo ⋅ [ht−1 , xt ] + bo ) (5)
1: Initialize: LSTM, QCBM, filter (Chemistry42)
2: Generate initial samples Xi from QCBM
3: while not converged do Update hidden state: ht = ot ⋅ tanh(Ct ) (6)
4:  Train LSTM with Xi
5:   LSTM generates a new compound from the current samples Xi Here, σ is the sigmoid activation function, W and b are the weight
6:  Validate the new compound with the filter matrices and bias vectors for each gate and xt is the input at time t69.
7:  if new compound is valid then Training an LSTM involves optimizing the network’s weights and
8:    Compute rewards for the new compound biases to minimize a specific loss function. This is typically accom-
9:    Compute probabilities P(Xi) for each new compound plished using gradient-based optimization algorithms such as sto-
10:    Train QCBM with Xi and P(Xi) chastic gradient descent or Adam70. The backpropagation through
11:     Generate new Xi from QCBM time algorithm was used to compute the gradients relative to the
12:  end if loss function, considering the sequential nature of the data71. The
13: end while networks were trained using the Adam optimizer with the NLL loss
function; to mitigate overfitting, regularization techniques such as
QCBM model. The QCBM model represents a quantum variational dropout were implemented72. The NLL loss for a single data point is
generative model, necessitating a classical optimizer to train its param- given by
eters. The total count of these parameters was computed with the
L( y, y)̂ = − log( yŷ ) (7)
number of qubits and layers defined in the model (Extended Data
Fig. 5). Upon specifying the parameters, denoted as θn, we obtain a
quantum state |ψ(θ)⟩. Here, each θn exerts an impact on the wave func- where y is the true class label and yŷ is the predicted probability for the
tion, expressed as ψ(θ). To optimize these parameters θ, the model is true class label y.

Nature Biotechnology
Brief Communication https://doi.org/10.1038/s41587-024-02526-3

The loss for a batch of data is the mean of the individual losses for distributions. A classical optimization algorithm recommends adjust-
each data point in the batch: ing parameters, which operates on the basis of the feedback received
from the evaluation of the circuit’s output. At each iteration, the quan-
N
1 tum circuit is sampled to produce a set of states. These states are then
ℒ=− ∑ log( yŷ i ) (8)
N i=1 compared against the target distribution and the difference between
them informs the direction and magnitude of parameter adjustments in
where N is the number of data points in the batch, yi is the true class the quantum circuit. This iterative process continues until the distribu-
label for the ith data point and yŷ i is the predicted probability for the tion generated by the QCBM closely aligns with the target distribution
true class label of the ith data point. or until a predefined convergence criterion is met.
In the hyperparameter tuning process, we used Optuna, an opti- In the context of QCBM training, the exact NLL functions as the
mization framework, to adjust parameters such as the number of hid- primary loss function, providing a quantitative measure of the differ-
den dimensions, embedding dimensions and layers within the model. ence between the distributions. The exact NLL for a QCBM is the nega-
The model presented in this research integrated a deep learning archi- tive sum of the logarithms of the probabilities that the quantum circuit
tecture. This architecture was designed to incorporate prior informa- assigns to the states in the training dataset. Mathematically, this is
tion (samples) into the generative process. Additionally, the model represented as NLL (θ) = −∑x∈D log pθ (x), where D is the set of data points
used Chemistry42 feedback in conjunction with QCBMs, aimed at and pθ(x) is the probability of observing state x under the current
enhancing its generative accuracy. Supplementary Figure 4 illustrates parameters θ of the quantum circuit. Minimizing the NLL involves
the proposed architecture at a cell level. The prior samples were com- adjusting θ such that the quantum circuit’s output distribution increas-
bined with input samples xt′ = X(i) ++xt in the LSTM cell. This combina- ingly resembles the empirical distribution of the data. This optimiza-
tion constituted two methods: adding and concatenating samples. The tion is typically carried out using gradient-based methods or other
LSTM’s operations were updated with the following operations: heuristic techniques suited to the quantum computing context. In our
project, we used COBYLA for our optimizer. As the NLL decreases, the
Prior sampling xt′ = X(i) ++xt OR xt′ = X(i) + xt (9)
fidelity of the QCBM in modeling the target distribution correspond-
ingly increases, indicating successful training of the quantum model.
Forget gate: ft = σ(Wf ⋅ [ht−1 , x′t ] + bf ) (10) Extended Data Fig. 5 shows the QCBM architecture and illustrates
its associated ansatz. We used linear topology for our project. Our
QCBM model was built with 16 qubits and four layers and we had 96
Input gate: it = σ(Wi ⋅ [ht−1 , x′t ] + bi ) (11)
parameters to optimize in total. The initial probability of the sam-
ples, P(X(i)), was computed on the basis of the rewards returned by
Candidate values: Ct̃ = tanh(WC ⋅ [ht−1 , x′t ] + bC ) (12) the Chemistry42 model. These reward-based probabilities were then
passed through a softmax function to ensure that they were normal-
ized and fell within the range of 0–1. The resulting values served as the
Update cell state: Ct = ft ⋅ Ct−1 + it ⋅ Ct̃ (13)
‘true’ probabilities of the samples and were used as the target values
during the model’s training process.
Output gate: ot = σ(Wo ⋅ [ht−1 , x′t ] + bo ) (14)
Algorithm 2. Algorithm 2 provides the pseudocode outlining the train-
ing regimen for the QCBM model. This process delineates the iterative
Update hidden state: ht = ot ⋅ tanh(Ct ) (15)
optimization of the QCBM parameters.

To generate samples, the process began with sampling from the 1: Initialize: QCBM model with a certain number of qubits and layers
prior, followed by the LSTM network processing these prior samples to 2: Set: Parameterized quantum state |ψ(θ)⟩
generate compounds representations. The compounds were validated 3: while not converged do
through the Chemistry42 platform, specifically tailored to assess ligand 4:  Compute exact NLL loss function
quality for the KRAS-G12D mutant. This methodology allows designing 5:  Compute gradient of exact NLL with respect to θ
ligands targeted at specific proteins. Moreover, the LSTM model is a 6:  Adjust parameters using an optimizer
classical approach for learning ligand structures and constructing a 7:  Validate the sample and compute its reward value
latent ligand space. The QCBM functions as a prior, guiding the LSTM 8:  if sample is valid then
in the generation of novel ligand samples. The procedure was subjected 9:    Compute rewards for the sample
to an iterative process to enhance the quality of ligands, which was 10:     Adjust probabilities P(Xi) on the basis of the rewards
evaluated using the Chemistry42 platform. 11:     Train QCBM model with adjusted probabilities
12:  end If
Quantum generative model: QCBM model. The QCBM is a variational 13: end while
quantum algorithm that uses the foundational principles of quantum
mechanics, particularly the Born rule, to generate complex and diverse Reporting summary
data samples. The core of our QCBM model is a parameterized quantum Further information on research design is available in the Nature
state |ψ(θ)⟩, where θ denotes the parameters or ansatz of our quantum Portfolio Reporting Summary linked to this article.
circuit. As per the Born rule, given a measurement basis, which is com-
monly the computational basis in our case, the probability of observing Data availability
a specific outcome |x⟩ is expressed as ∣〈x∣ψ(θ)〉∣2. The datasets generated and analyzed during the current study, along
Training a QCBM involves optimizing the parameters of the with the generative model, are available via Zenodo at https://doi.org/
quantum circuit to produce a probability distribution that closely 10.5281/zenodo.11137638 (ref. 73).
approximates the target distribution (probability computed by Chem-
istry42 reward values). This process is fundamentally iterative, where Code availability
the quantum circuit parameters, denoted as θ, are adjusted in each The full code for running our generative model is available via Zenodo
step to reduce the discrepancy between the generated and target at https://doi.org/10.5281/zenodo.11137638 (ref. 73).

Nature Biotechnology
Brief Communication https://doi.org/10.1038/s41587-024-02526-3

References 57. Williamson, M. P. Using chemical shift perturbation to

37. Olivecrona, M., Blaschke, T., Engkvist, O. & Chen, H. Molecular characterise ligand binding. Prog. Nucl. Magn. Reson. Spectrosc.
de-novo design through deep reinforcement learning. 73, 1–16 (2013).
J. Cheminform. 9, 48 (2017). 58. Zagribelnyy, B., Putin, E. O., Fedorchenko, S A., Ivanenkov, Y. A.
38. Bombarelli, R. G. et al. Automatic chemical design using & Zavoronkovs, A. Retrosynthesis-related synthetic accessibility.
a data-driven continuous representation of molecules. WO patent 2021/229454 A1 (2021).
ACS Cent. Sci. 4, 268–276 (2018). 59. Welsch, M. et al. Multivalent small-molecule pan-RAS inhibitors.
39. Krenn, M. et al. SELFIES and the future of molecular string Cell 168, 878–889 (2017).
representations. Patterns 3, 100588 (2022). 60. Kwan, A. K., Piazza, G. A., Keeton, A. B. & Leite, C. A. The path to
40. Zang, C. & Wang, F. Moflow: an invertible flow model for the clinic: a comprehensive review on direct KRasG12C inhibitors. J.
generating molecular graphs. In Proc. of the 26th ACM SIGKDD Exp. Clin. Cancer Res. 41, 27 (2022).
International Conference on Knowledge Discovery and Data Mining 61. Parikh, K. et al. Drugging KRas: current perspectives and
(eds Gupta, R. & Liu, Y.) 617–626 (ACM, 2020). state-of-art review. J. Hematol. Oncol. 15, 152 (2022).
41. Segler, M., Kogej, T., Tyrchan, C. & Waller, M. P. Generating 62. Srisongkram, T., Khamtang, P. & Weerapreeyakul, N. Prediction
focused molecule libraries for drug discovery with recurrent of KRasG12C inhibitors using conjoint fingerprint and machine
neural networks. ACS Cent. Sci. 4, 120–131 (2018). learning-based QSAR models. J. Mol. Graph. Model. 120, 108197
42. Brown, N., Fiscato, M., Segler, M. & Vaucher, A. C. Guacamol: (2023).
benchmarking models for de novo molecular design. J. Chem. Inf. 63. Nagasaka, M. et al. KRas G12C Game of Yhrones, which direct
Model. 59, 1096–1108 (2019). KRas inhibitor will claim the iron throne? Cancer Treat. Rev. 84,
43. Jensen, J. H. A graph-based genetic algorithm and generative 101974 (2020).
model/Monte Carlo tree search for the exploration of chemical 64. Grygorenko, O. O. et al. Generating multibillion chemical space
space. Chem. Sci. 10, 3567–3572 (2019). of readily accessible screening compounds. iScience 23, 101681
44. Nigam, A. K., Pollice, R. & Guzik, A. A. Parallel tempered genetic (2020).
algorithm guided by deep neural networks for inverse molecular 65. Krenn, M., Häse, F., Nigam, A. K. & Friederich, P. Self-referencing
design. Digit. Discov. 1, 390–404 (2022). embedded strings (SELFIES): a 100% robust molecular string
45. Frandsen, A. et al. Tyr702 is an important determinant of agonist representation. Mach. Learn. Sci. Technol. 1, 045024 (2020).
binding and domain closure of the ligand-binding core of GLUR2. 66. Lo, A. et al. Recent advances in the self-referencing embedded
Mol. Pharmacol. 67, 703–713 (2005). strings (SELFIES) library. Digit. Discov. 2, 897–908 (2023).
46. Zhang, L. et al. Crystal structure of SARS-CoV-2 main protease 67. Hochreiter, S. & Schmidhuber, J. Long short-term memory. Neural
provides a basis for design of improved α-ketoamide inhibitors. Comput. 9, 1735–1780 (1997).
Science 368, 409–412 (2020). 68. Gers, F. A., Schmidhuber, J. & Cummins, F. Learning to forget:
47. Ring, A. M. et al. Adrenaline-activated structure of continual prediction with LSTM. In Proc. of the 9th International
β2-adrenoceptor stabilized by an engineered nanobody. Nature Conference on Artificial Neural Networks (eds Sandoval, F. et al.)
502, 575–579 (2013). 850–855 (Springer, 1999).
48. Alhossary, A., Handoko, S. D., Mu, Y. & Kwoh, C. K. Fast, accurate, 69. Gers, F. A. & Schmidhuber, J. Recurrent nets that time and count.
and reliable molecular docking with QuickVina 2. Bioinformatics In Proc. of the IEEE–INNS–ENNS International Joint Conference on
31, 2214–2216 (2015). Neural Networks (eds Amari, S.-I. et al.) 189–194 (IEEE, 2000).
49. Koes, D. R., Baumgartner, M. P. & Camacho, C. J. Lessons 70. Kingma, D. P. & Ba, J. Adam: a method for stochastic optimization.
learned in empirical scoring with smina from the CSAR 2011 In Proc. of the 3rd International Conference on Learning
benchmarking exercise. J. Chem. Inf. Model. 53, 1893–1904 Representations (ed. Yue, X.-G.) 1–13 (ACM, 2015).
(2013). 71. Werbos, P. J. Backpropagation through time: what it does and how
50. Voigt, J. H., Bienfait, B., Wang, S. & Nicklaus, M. C. Comparison to do it. Proc. IEEE 78, 1550–1560 (1990).
of the NCI open database with seven large chemical structural 72. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I. &
databases. J. Chem. Inf. Comput. Sci. 41, 702–712 (2001). Salakhutdinov, R. Dropout: a simple way to prevent neural
51. Ihlenfeldt, W. D., Voigt, J. H., Bienfait, B., Oellien, F. & Nicklaus, M. networks from overfitting. J. Mach. Learn. Res. 15, 1929–1958
C. Enhanced CACTVS browser of the open NCI database. J. Chem. (2014).
Inf. Comput. Sci. 42, 46–57 (2002). 73. Vakili, M. G. Quantum computing-enhanced algorithm unveils
52. Akiba, T., Sano, S., Yanase, T., Ohta, T. & Koyama, M. Optuna: a novel inhibitors for KRas. Zenodo https://doi.org/10.5281/
next-generation hyperparameter optimization framework. In Proc. zenodo.11137638 (2024).
of the 25th ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining (eds Teredesai, A. & Kumar, V.) 2623– Acknowledgements
2631 (ACM, 2019). We want to thank A. Perdomo-Ortiz, M. Mauri, B. Dellabetta, V.
53. Lim, S. M. et al. Therapeutic targeting of oncogenic K-Ras by a Vargas-Calderón, A. Cheng, J. Miller, M. Bagherimehrab and K. Nakaji
covalent catalytic site inhibitor. Angew. Chem. Int. Ed. Engl. 53, for their valuable discussion and support in this research. We thank H.
199–204 (2014). Arthanari for providing us with his NMR resources. Additionally, we are
54. Pálfy, G., Vida, I. & Perczel, A. 1H, 15N backbone assignment and thankful to the Defense Advanced Research Projects Agency for their
comparative analysis of the wild type and G12C, G12D, G12V funding (grant number HR0011-23-3-0017), which greatly supports
mutants of K-Ras bound to GDP at physiological pH. Biomol. NMR our scientific endeavors. A.K.N. acknowledges funding from the
Assign. 14, 1–7 (2020). Bio-X Stanford Interdisciplinary Graduate Fellowship. A.A.-G. thanks
55. Delaglio, F. et al. NMRPipe: a multidimensional spectral A. G. Frøseth for his generous support. A.A.-G. also acknowledges
processing system based on unix pipes. J. Biomol. NMR 6, the generous support of Natural Resources Canada and the Canada
277–293 (1995). 150 Research Chairs program. C.G. was funded through the Blue Sky
56. Vranken, W. et al. The CCPN data model for NMR spectroscopy: Kinase Project at St. Jude Children’s Research Hospital. The research
development of a software pipeline. Proteins 59, 687–696 in the I.S. lab is supported by the grants from the Canadian Institute for
(2005). Health Research, Genome Canada, Ontario Research Fund, Ontario

Nature Biotechnology
Brief Communication https://doi.org/10.1038/s41587-024-02526-3

Genomics Institute, Cystic Fibrosis Canada and Toronto Innovation is a cofounder and consultant of Virtual Discovery and Quantum
Acceleration Partners. This research was undertaken thanks partly to Therapeutics. The remaining authors declare no competing interests.
funding provided to the Acceleration Consortium of the University of
Toronto from the Canada First Research Excellence Fund. Additional information
Extended data is available for this paper at
Author contributions https://doi.org/10.1038/s41587-024-02526-3.
M.G.V. led the project, developed the quantum machine learning
algorithm, benchmarked the model and contributed substantially Supplementary information The online version contains supplementary
to writing the manuscript. C.G. led the virtual screening and NMR material available at https://doi.org/10.1038/s41587-024-02526-3.
experiments. A.K.N. worked on the genetic algorithm and generated
the initial data. The Insilico team focused on synthesizing the Correspondence and requests for materials should be addressed
compounds and performing the SPR assays. J.S. led the cell-based to Christoph Gorgulla, AkshatKumar Nigam, Igor Stagljar,
assays conducted in the I.S. lab. The remaining authors worked in Alán Aspuru-Guzik or Alex Zhavoronkov.
parallel to validate the findings, discuss the results and contribute to
the preparation and analysis of the final results. Peer review information Nature Biotechnology thanks Su-Yang Xu and
the other, anonymous, reviewer(s) for their contribution to the peer
Competing interests review of this work.
A.A.G. serves as the chief visionary officer and is a board member of
Kebotix. A.Z., A.A., D.B., D.P., X.D., J.L., E.R., F.R. and F.M. are affiliated Reprints and permissions information is available at
with Insilico Medicine. Y.C. and D.V. are affiliated with Zapata AI. C.G. www.nature.com/reprints.

Nature Biotechnology
Brief Communication https://doi.org/10.1038/s41587-024-02526-3

Extended Data Fig. 1 | Comparative Benchmarking of Quantum and Classical approaches with varied priors. The performance of the Quantum Circuit Born
Ligand Design Methods. (A, B) Comparative analysis of our hybrid approaches Machine (QCBM) was assessed using both a quantum simulator (Sim) and a
with varied priors. The performance of the Quantum Circuit Born Machine hardware backend (HW), and contrasted with a Multi-bases QCBM (MQCBM)
(QCBM) was assessed using both a quantum simulator (Sim) and a hardware operating solely on a quantum simulator (Sim), as well as an LSTM model devoid
backend (HW), and contrasted with a Multi-bases QCBM (MQCBM) operating of quantum priors (representing a fully classical architecture). We calculated the
solely on a quantum simulator (Sim), as well as an LSTM model devoid of number of generated molecules that met a series of synthesizability and stability
quantum priors (representing a fully classical architecture). We calculated criteria as stipulated by the Tartarus benchmarking platform (referred to as Local
the number of generated molecules that met a series of synthesizability and Filters)26 and by Chemistry42 (referred to as Chemistry42 Filters). To generate
stability criteria as stipulated by the Tartarus benchmarking platform (referred Figure A, we repeated the experiments five times, sampling n=1,000 compounds
to as Local Filters)26 and by Chemistry42 (referred to as Chemistry42 Filters). To in each repetition and applying the filter. The reported values represent the
generate Figure A, we repeated the experiments five times, sampling n=1,000 mean for each data point, with error bars indicating the standard deviation
compounds in each repetition and applying the filter. The reported values across the repetitions (Mean ± Std.). more detailed is reported in Supplementary
represent the mean for each data point, with error bars indicating the standard Information Table S3.2. (C) Success rate of generating molecules that meet
deviation across the repetitions (Mean ± Std.). more detailed is reported in Tartarus’s filter criteria as a function of the number of qubits used in modeling
Supplementary Information Table S3.2. (C) Success rate of generating molecules priors for the QCBM. We repeated the experiments five times, sampling n=5,000
that meet Tartarus’s filter criteria as a function of the number of qubits used in compounds in each repetition and applying the filter. The reported values
modeling priors for the QCBM. We Comparative Benchmarking of Quantum represent the mean for each data point, with error bars indicating the standard
and Classical Ligand Design Methods. (A, B) Comparative analysis of our hybrid deviation across the repetitions.

Nature Biotechnology
Brief Communication https://doi.org/10.1038/s41587-024-02526-3

Extended Data Fig. 2 | Protein-detected NMR experiments. Protein-detected are mapped onto the G12D structure and displayed in a color gradient. Residues
NMR experiments elucidate the binding mode of Compound 18-2 to KRAS-G12D. that could not be assigned are shown in grey. The majority of significant CSPs are
(A) Significant chemical shift perturbations (CSPs) and intensity changes due observed near the Switch-II pocket; however, CSPs are also noted in the α-1 helix,
to chemical exchange are observed upon binding to the compound. Residues Switch-I, and α-4 regions, possibly indicating conformational changes upon
lining the Switch-II pocket that show significant CSPs are highlighted in the inset, binding to the compound. (C) Zoomed-in region of the protein, highlighting
including Gly10 (from the P-loop), Phe78 and Gly77 (from the α-2 helix adjacent residues from the Switch-II pocket that exhibit CSPs.
to the Switch-II region), Ser89, and Asp92 (from the β-4 sheet). (B) The CSP values

Nature Biotechnology
Brief Communication https://doi.org/10.1038/s41587-024-02526-3

Extended Data Fig. 3 | Quantum-Enhanced Generative Model for Drug encoding) and quantum data are merged by addition or concatenation. The
Discovery Applications. (B) Hybrid model combining a Quantum Circuit Born resultant samples, X’(t), are then input to the LSTM cell. (C) Quantum prior
Machine (QCBM) with Long Short-Term Memory (LSTM). This model iteratively component described as a QCBM, generating samples from quantum hardware
trains using prior samples from quantum hardware. (A) Integration method of each training epoch and trains with a reward value, P (x)=Softmax(R(x)),
prior samples into the LSTM architecture. Molecular information (in SELFIES calculated using Chemistry42 or a local filter.

Nature Biotechnology
Brief Communication https://doi.org/10.1038/s41587-024-02526-3

Extended Data Fig. 4 | Pharmacophore Model Depiction. Pharmacophore supports structural integrity, a green sphere highlights a hydrophobic moiety
Model Depiction for KRAS Inhibitor TH-Z816 Based on the Co-crystallized Ligand essential for binding affinity, and a cyan sphere indicates a hydrogen bond donor
Structure Analyzed with Chemistry42 (PDB: 7EW9). This figure illustrates the that contributes to interaction specificity with the KRAS protein. The protein
key pharmacophoric features identified from the ligand structure of TH-Z816 structure is shown on the right, while the pharmacophore interactions within the
bound to KRAS G12D. A blue sphere represents a critical ring system that KRAS Switch-II binding pocket are detailed on the left.

Nature Biotechnology
Brief Communication https://doi.org/10.1038/s41587-024-02526-3

Extended Data Fig. 5 | Quantum Circuit Born Machine (QCBM) Model. gates, including parameterized rotations (Rx, Rz) and entangling CNOT gates, are
Schematic representation of the Quantum Circuit Born Machine (QCBM) orchestrated to evolve the initial state |0⟩ into a complex quantum state |Ψ(0)⟩.
implemented in our numerical experiments, illustrating a variational quantum The outcome is measured, and the resulting data are used by the classical
circuit with a configuration of three layers and four qubits. In practice, our optimizer to iteratively refine the parameters θ, thus leading the circuit towards
numerical experiments utilized a system with 16 qubits. The depicted quantum an optimal solution for ligand generation.

Nature Biotechnology
μ μ
₂