Insider Brief
- Quantum annealing–based drug design, demonstrated by PolarisQB’s QuADD platform running on a D-Wave Advantage system, can generate and optimize drug-like molecular candidates in minutes to hours rather than weeks or months, significantly reducing early-stage discovery time and cost.
- In a head-to-head study using Thrombin as a test case, QuADD produced higher-quality, more synthesizable leads with stronger predicted binding affinities and better drug-like properties than a representative generative AI diffusion model, while requiring roughly 30 minutes of computation versus about 40 hours.
- By framing molecular discovery as a constrained combinatorial optimization problem, annealing quantum computers prioritize viable, drug-ready candidates over sheer molecular diversity, improving hit-to-lead efficiency and lowering downstream experimental attrition.
- Photo by Pavel Danilyuk on Pexels
PRESS RELEASE — In a drug design campaign, researchers look for molecules with a specific set of properties, including biological activity, low toxicity, metabolic stability, solubility, etc., all specifically selected to create the ideal candidate for delivering more effective treatments and cures.
Quantum computing, and specifically quantum annealing, now shows the potential to significantly accelerate what could previously take years, with the possibility of delivering new compounds that meet these desired characteristics in weeks or months, and reducing the labor and laboratory costs of early-stage drug discovery.
Annealing quantum computers excel at combinatorial optimization, making them well-suited for supercharging molecular design, screening, and optimization workflows, including tasks such as selecting the optimal fragments used to build molecular candidates, designing core structures for these molecules, or deciding which molecules to substitute in or out from enormous chemical libraries. These capabilities transform drug discovery into a tailorable design problem.
To take advantage of quantum’s unique capabilities to advance drug design, Polaris Quantum Biotech (PolarisQB) built Quantum-Aided Drug Design (QuADD) – the first drug discovery software platform built around quantum computing for lead identification. The calculations used to generate the molecular candidates are run on a D-Wave Advantage system with over 5000 qubits.
By rapidly prioritizing synthesizable, drug-like leads, the QuADD platform can shorten the hit-to-lead and lead-optimization stages, freeing experimental resources to focus on a smaller, higher-quality set of candidates. Generative AI and other classical alternatives to the QuADD platform require weeks or months of computational effort to deliver results comparable to, or worse than, those produced by PolarisQB’s platform in a few short hours.
A new paradigm for drug discovery
Leveraging an annealing quantum computer, QuADD can explore a chemical space up to about 10^30 theoretical molecules while optimizing for properties such as permeability, solubility, toxicity, metabolic stability, and binding pocket fit in just a few hours.
Using QuADD, annealing quantum computers generate low-energy, high-quality solutions, such as molecules that satisfy many property constraints, improving both the quality and diversity of generated drug-like structures (molecules with properties suitable for medicines). Classical computing resources can further build upon the QuADD-generated structures, such as using AI models to evaluate toxicity and molecular mechanics to refine them further.
“PolarisQB recently completed an objective comparison of its platform to non-quantum-based methods, and the improvements were impressive,” said Dr. Ian Reynolds, President at YaghPenn Consulting. “This platform is ready to take off and have a big impact on the drug discovery industry.”
Our latest study (preprint on arXiv, currently undergoing peer review) explores QuADD’s capabilities. It compares QuADD with a broadly representative AI-based drug design system. QuADD uses quantum computing to generate drug-like molecules optimized for specific binding sites. For comparison in this study, Generative AI models, such as Bond and Interaction generating Diffusion model (BInD), use AI-based diffusion-based algorithms to explore chemical space and design new molecules.
Both systems were evaluated for their ability to generate novel, structurally diverse, and biologically relevant molecules within a defined protein-binding site, using the known case study of Thrombin, an enzyme that enables blood clotting.
For both systems, basic default configurations were used to design over 3,000 molecules as the test data for comparison. Then, as is usually done during a drug-design campaign, we selected the 100 best lead compounds from each system’s molecules, based on predicted binding affinity, and compared their properties (Figure A).
In evaluating these systems, QuADD required approximately 30 minutes to generate 3,000 molecules from which we selected 100 molecular candidates designed to bind to Thrombin. In contrast, BInD, the Generative AI model, required approximately 40 hours to generate the same number of molecules on a node equipped with a single NVIDIA GPU.
Quality over quantity in molecular design
Drug design aims not only to determine binding affinity for the target binding pocket but also to assess a range of other properties, since the project goal is to identify drug molecules rather than tool compounds used solely for research. This is where the strength of annealing quantum computing truly shines, as it solves Quadratic Unconstrained Binary Optimization (QUBO) or multi-objective optimization problems.
Specifically, for the QuADD platform, we adapted the Knapsack Problem – a well-known combinatorial optimization problem – to handle molecular fragments that are the best fits for subregions of the binding site. These fragments must also comply with a set of retrosynthetic rules (which guide drug development by working backwards from target molecules). At the same time, the objective (binding affinity) is achieved under constraints such as one fragment per region, several drug-like properties, and fragments compatibility.
When compared, both methods produced novel compounds that fit the target binding site. However, while BInD generated a more diverse set of molecules, QuADD’s outputs showed superior binding affinities, stronger interaction fidelity, and overall better drug-like characteristics.
By confining the search to a chemically meaningful yet bounded subspace, QuADD ensures that results are not only innovative but also actionable – meaning they are highly likely to be synthesizable and suitable for experimental testing and confirmation.
In contrast, BInD prioritizes creative generation, using probabilistic AI-based models to move beyond established chemical frameworks. While this strategy can yield novel and highly diverse scaffolds, it also introduces considerable risk: many candidates may be synthetically infeasible, pharmacologically irrelevant, or unlikely to bind effectively to the target protein. These factors are likely to result in high attrition rates and require more substantial downstream filtering and validation. AI methods may be an incremental improvement over traditional methods, but they are still bound to fundamental constraints of classical computing systems.
To describe our results in a finer detail, we compared several molecularly relevant properties between the two sets of molecules. We sought to determine which system provided the superior optimization of physicochemical and pharmacophoric characteristics, both specific to the target environment and relevant to drug-like properties.
For example, synthetic complexity represents the steps (time and materials) required to synthesize a molecule. This metric appraises the putative costs of structures generated from both QuADD and BInD. Inevitably, fewer synthetic steps per compound result in significantly lower synthesis costs and a more rapid turnaround upon ordering. When considering a purchase of 100 compounds, even at milligram yields, additional steps per compound can increase synthesis costs by an order of magnitude or more. We have found that QuADD structures have a lower synthetic complexity, making them more realistic as drug candidates. This was evident for other properties, such as drug-like property scores (Molecular weight and Topological Polar Surface Area), and binding site interactions.
While many metrics can be used to select a final set of candidates for synthesis in a drug discovery campaign, it is standard practice to use an estimate of binding affinity in the selection process (the higher the binding affinity, the more effective the drug). In Figure B, we see that the predicted binding affinities of the top 100 QuADD-generated molecules have substantially stronger predicted binding affinities than those of the BInD-generated molecules. Lower (more negative) scores indicate a stronger predicted binding affinity. There are a small number of molecules whose binding affinities overlap well with the QuADD distribution. Notably, this overlap occurs at the best binders from BInd and the weakest binders from QuADD.
Figure B shows that, on average, the QuADD platform improves the predicted binding affinity of drug candidates by (at least) 1kcal/mol, which is an order of magnitude better predicted binding.
Faster discovery, superior drugs
While the current AI trend produces vast numbers of molecules, unfortunately, these molecules often fail basic drug-likeness or synthetic tests. At the same time, QuADD’s quantum optimization consistently generates real, synthesizable candidates with superior binding and pharmacological properties.
Our study highlights trade-offs which can significantly impact a campaign’s budget and timeline: while BInD’s candidates exhibit greater structural diversity, many of them fail essential filters for drug-likeness, binding affinity, or synthetic viability. QuADD, in contrast, delivers candidates that reliably meet stringent pharmaceutical criteria – routinely passing drug-likeness screens, recapitulating key protein-ligand interactions, and achieving stronger predicted affinities. QuADD produces a high rate of viable molecules in a significantly faster timeframe at a lower cost than BInD.
These findings demonstrate the advantages of quantum computing for optimizing molecular design and reinforce PolarisQB’s commitment to harnessing cutting-edge quantum technologies to accelerate drug discovery. QuADD’s superior performance positions it as a transformative platform for developing next-generation therapeutics.
Auransa, a clinical-stage, AI-driven biopharma company, is working with PolarisQB to implement these transformative capabilities in drug discovery today.
“The QuADD platform was valuable in helping us design and prioritize compounds with targeted physicochemical properties beyond biological activity, such as better fit to challenging binding pockets,” said Pek Lum, founder and CEO of Auransa. “Its ability to rapidly explore and triage a very large chemical design space helped us move from ideas to testable compounds rapidly and more efficiently and accelerated our learning cycle.”
The substantially shorter time required by QuADD to generate a candidate pool of molecules allows researchers to modify and refine search and selection criteria for follow-up runs.
QuADD’s constraint-driven optimization provides a unique combination of reliability, efficiency, and alignment with drug development needs. It stands out as a powerful platform for generating not only novel molecules but also truly translatable, drug-like candidates, delivering a practical advantage in the early stages of modern drug discovery. While we’ve always believed in the power of quantum technology and in its implementation in QuADD, the results now clearly show quantum’s edge.
