Researchers Develop Breakthrough Quantum AI Tool that Unlocks ‘Undruggable’ Cancer Target
Researchers from Insilico Medicine and the Acceleration Consortium at the University of Toronto (CA, U.S.) have developed a quantum-classical AI pipeline that can design new drugs against a major cancer-related protein. This opens the door to a cancer target previously considered undruggable.
Why Can’t Scientists Target One of the Most Common Mutations in Cancer?
KRAS is a protein that plays a crucial part in human biology; it acts as a molecular ‘switch’ that orchestrates cell growth and apoptosis at exactly the right times. Its critical role means that mutations affecting its function or shape can disrupt this balance, triggering uncontrolled cell division that may drive cancer formation.
The gene that encodes the KRAS protein is one of the most frequently mutated genes seen in cancer patients. While its prevalence varies across subtypes, these mutations are found in up to 90% of pancreatic cancer cases.
This has made KRAS a particularly attractive target for scientists in cancer research; targeting this protein would inhibit cellular division and promote apoptosis, leading to the destruction of tumor cells.
For decades, KRAS has puzzled researchers as an undruggable target. Usually, drugs are engineered to bind to unique crevices or pockets on the surface of target proteins. However, the surface of KRAS is relatively smooth, meaning it is hard to find compounds that ‘stick.’ While there has been some success in this area, progress is slow and strategies for finding new KRAS-targeting drugs are limited.
One promising strategy is the use of computational tools to find new compounds. Indeed, with the use of machine learning techniques increasing, researchers are looking to harness the power of AI to streamline drug development.
A new study published in Nature marks the first of its kind, using a quantum-classical hybrid model to predict inhibitors for KRAS.
“As many as 85% of all human proteins are thought to be ‘undruggable,’” says Says Alex Zhavoronkov, CEO of Insilico Medicine and co-author of the study. “This is a major challenge facing the development of new cancer treatments, and one that AI is uniquely positioned to help.”
Combining the Best of Both Worlds Into One State-of-the-art Workflow
But what exactly is a quantum-classical hybrid model, and why is this so exciting?
This model synthesizes two core frameworks: Quantum Circuit Born Machines (QCBMs) and Long Short-Term Memory (LSTMs) generative models, forming the ‘quantum’ and ‘classical’ components of the name, respectively. The workflow also incorporates Insilico medicine’s generative AI platform Chemistry42 to validate the predictions of the LSTM and QCBM models.
An LSTM is a type of recurrent neural network that generates new data sequences by learning underlying patterns. In this case, the LSTM learns the chemical structures of KRAS inhibitors as character strings and generates new potential candidates.
Although classical AI systems have contributed majorly to the field of drug discovery, they are restricted when it comes to exploring large chemical spaces, i.e., finding all the targetable parts of proteins. These models also tend to approximate quantum behaviors, meaning they neglect the precise interactions between sub-atomic particles, such as electrons.
On the other hand, quantum computational methods allow researchers to model intricate molecular details, such as superposition and entanglement. By using complex probability distributions to learn and predict high-dimensional data, QCBMs are extraordinarily powerful tools, particularly when looking at large biological targets like proteins.
Nonetheless, QCBMs are computationally expensive, difficult to train, and can be highly sensitive to noise or disturbances in data .
Combining quantum and classical designs therefore overcomes each system’s limitations:
“While we often hear of the potential synergy between artificial intelligence and quantum computing, researchers have now brought this dream to life and have developed a tool to fast-track the measurement of protein interactions for drug discovery,” Says Zhavoronkov.
From an initial dataset of 1.1 million molecules, the AI model selected 15 of the most-promising drug candidates with potential KRAS-inhibitory activity. The group synthesized these candidates and tested their biological activities, where two lead compounds demonstrated impressive KRAS inhibition. In particular, one compound maintained robust inhibition across different mutation subtypes, which may have even affected the protein differently.
Having showcased the potential of quantum-classical hybrid AI methods to find KRAS inhibitors, the team is now turning their attention to new compounds for other traditionally undruggable targets. They also hope to advance the two lead compounds –identified in this study– through further testing and optimization using their hybrid model.
Will this Change the Trajectory of Drug Discovery?
This landmark paper has set the scene for the future of AI-driven drug discovery. The researchers have leveraged the strenghts of both quantum and classical methods to effectively design two new potential drugs that both target a notoriously undruggable protein.
Potentially, this breakthrough may now open the door to addressing other targets that scientists have long considered untouchable, shining a spotlight on unexplored areas of research.
Adjacent to this, utilizing powerful AI models –such as the one in this study– could drastically cut down the time spent on laborious drug discovery processes, helping to bring more effective drugs into clinical spaces on a signifciantly reduced timescale.
Igor Stagljar, co-investigator of the study, concludes:
“With computational approaches like this, we have the potential to shorten the preclinical phase of drug discovery by years.”