Quantum computing + AI, breaking the bottleneck in drug development?

KRAS is an extremely common "problematic protein" in cancer, present in approximately one-quarter of human tumors with KRAS mutations. This mutation allows cells to proliferate uncontrollably, leading to cancer. Although these mutations are very common and pose significant harm, there are currently only two FDA-approved drugs specifically targeting KRAS mutations, and they can only extend patients' survival to a limited extent. There is an urgent need for new KRAS therapies that can bring greater benefits to many cancer patients.

In a recent study published in Nature Biotechnology, a cross-disciplinary collaboration involving Insilico Medicine, the University of Toronto in Canada, and other research institutions such as St. Jude Children's Research Hospital, attempted to design new inhibitor molecules against "undruggable" KRAS from scratch using "quantum computing + classical computing + generative AI."

This study demonstrates for the first time the potential advantages of combining quantum computing with AI in the early drug discovery process, bringing new hope for treatment options targeting difficult-to-address targets.

Quantum Computing + AI: How to Construct the Drug Molecule Generation Process

The research team proposed a quantum-classical hybrid generative framework: combining Quantum Circuit Born Machine (QCBM) and Long Short-Term Memory (LSTM) networks to collaboratively design new molecules. Specifically, they trained the quantum-classical hybrid model using a custom dataset containing 1.1 million molecules. This vast data source includes:

• 650 molecules that have been confirmed in the literature to block KRAS,
• 850,000 analogs derived from known KRAS inhibitors using the STONED-SELFIES algorithm,
• 250,000 molecules obtained through the virtual screening platform VirtualFlow.

Such rich training data allows the quantum-classical hybrid model to learn a broader "chemical space," laying the foundation for generating diverse candidate molecules.

Next comes the collaborative generation process of quantum and classical.

In this quantum-classical hybrid model,

• QCBM: Acts as the quantum generative model, using quantum circuits to learn complex probability distributions and generate molecular structures that are similar to but "not yet explored" in the training data. It also serves as a "prior," guiding the LSTM's molecular sequence generation.
• LSTM: Leverages the advantages of classical AI models, capable of understanding and generating sequential data. By incorporating the probability distribution output from QCBM, LSTM can more accurately grasp molecular diversity when generating new chemical structures, avoiding overfitting or converging on familiar structures

In practical applications, the research team first generated 1 million candidate molecules at once using a hybrid model. Then, they systematically evaluated and screened these molecules using the generative artificial intelligence engine Chemistry42 developed by InSilico Medicine, assessing multiple dimensions including drug-like properties, docking scores, and synthetic accessibility, ultimately selecting 15 of the most promising candidate molecules for laboratory testing.

From Cloud Screening to Experimental Validation

Compared to traditional drug discovery, this method does not rely on large-scale physical compound libraries for expensive and lengthy high-throughput screening. Instead, most of the screening work can be done in the cloud, significantly reducing costs and time. In the final laboratory phase, "wet lab" tests were conducted on the 15 selected molecules, resulting in two particularly outstanding candidates.

One of the molecules, named ISM061-018-2, exhibits strong targeting activity against KRAS and shows no significant cytotoxicity. Additionally, it has inhibitory activity against wild-type KRAS and various common mutant KRAS types (as well as wild-type HRAS and NRAS), demonstrating potential as a "pan-RAS inhibitor."

The other molecule, ISM061-022, shows more efficient inhibitory effects against certain mutant KRAS types (such as G12R and Q61H) and is also expected to develop into a candidate for a broad-spectrum anticancer drug.

It is worth noting that current research has not yet proven that this quantum-classical hybrid approach is "superior" to purely classical methods, but it at least indicates the feasibility and potential acceleration of quantum computing in early drug discovery. As quantum computing hardware continues to upgrade, its application prospects in generative models will also expand accordingly.

Dr. Alán Aspuru-Guzik, a professor of chemistry and computer science at the University of Toronto, stated: "This is a proof-of-principle study that preliminarily shows that quantum computers can be integrated into modern AI-driven drug development processes and successfully design active molecules that can bind to biological targets. Although we have not yet seen 'absolute advantages of quantum computing over classical methods,' we expect that as quantum hardware capabilities improve, related algorithms will become increasingly 'impressive.'"

After achieving early success against KRAS, the research team plans to extend this quantum-classical hybrid model to more "undruggable" protein targets. Similar to KRAS, these proteins are small and lack "pockets" on their surfaces that can stably bind to compounds, making them some of the most challenging targets in drug development. Researchers will also continue to optimize the KRAS lead compounds obtained and validate them in animal models, striving to bring more effective new-generation molecules to cancer patients.

Dr. Alex Zhavoronkov, founder and CEO of InSilico Medicine, stated: "Up to 85% of human proteins are considered 'undruggable,' and how to 'unlock possibilities' from these proteins has always been a challenge in cancer research Artificial intelligence can uniquely demonstrate its power in tackling this difficult challenge. We are very pleased to collaborate with the University of Toronto to integrate quantum computing into AI-driven drug discovery processes, seeking more possibilities for human health."

The Prospects of Integrating Quantum and AI are Promising

In fact, this is not the first collaboration between InSilico Medicine and the University of Toronto. As early as 2023, they published their first joint paper in the Journal of Chemical Information and Modeling, gradually replacing different parts of the classical generative model MolGAN with Variational Quantum Circuits (VQC) through multiple experimental scenarios, exploring the application of quantum generative adversarial networks in small molecule drug discovery.

The latest results published in Nature Biotechnology once again confirm the potential value of quantum computing in the drug design phase. With the further integration of quantum computing technology and AI generative models, it may be possible to more quickly and accurately screen active molecules targeting those "difficult targets," bringing hope to more patients in the future.

Although it cannot yet be asserted that quantum computing has surpassed classical algorithms, InSilico Medicine, as a pioneer in the AI pharmaceutical field, is actively seeking ways to combine quantum computing with AI to gain a first-mover advantage when breakthroughs in quantum computing technology occur. This idea also resonates with the optimistic expectations for AI drug development internationally.

Nobel laureate and Google DeepMind CEO Demis Hassabis recently stated at the World Economic Forum in Davos that drugs designed based on AI are expected to enter clinical trials by the end of this year.

These drugs are being developed by Isomorphic Labs, a subsidiary of Alphabet, aiming to reshape the drug discovery process from first principles. Hassabis pointed out that AlphaFold technology has successfully predicted the structures of 200 million proteins, providing unprecedented possibilities for precise research and development.

Against this backdrop, more research institutions, startups, and large multinational pharmaceutical companies are continuously exploring the combination of AI with cutting-edge technologies such as quantum computing. AI can not only help scientists quickly screen potential candidates from a vast molecular space but also further leverage the powerful computing capabilities of quantum computing to find new molecular design ideas that better meet the needs of "undruggable" targets