Qubit Pharmaceuticals, a deeptech company specialising in the discovery of new drug candidates, unveils the "world’s most advanced" quantum AI model to unlock an entirely new range of therapeutics.
Shutterstock AI Shutterstock
The quantum AI model, developed in partnership with Sorbonne University, is capable of modelling and simulating the behaviour of molecules with a level of precision and computational speed never before achieved. The accuracy will massively reduce the costly stage of laboratory experimentation in the development of new molecules by replacing the chemical synthesis of drug candidates.
“This simulation method would reduce the cost of the drug discovery phase enormously,” said Jean-Philip Piquemal, Professor at Sorbonne University and director of the Theoretical Chemistry Laboratory (Sorbonne University/CNRS), co-founder and scientific director of Qubit Pharmaceuticals. “The model is as accurate as the experiment; we can generate a huge number of new ideas, failing quickly and cheaply ‘in silico’ before moving on to laboratory testing with molecules that have passed the tests with flying colours.”
The team used computing power from GENCI, EuroHPC and Argonne to create FeNNix-Bio1, a foundational model built on millions of meticulous molecular simulations. It was trained on the world's most accurate molecular chemistry database, simulated to the highest possible chemical precision. By training on these elementary bricks, the foundation model learns the laws of chemistry and physics, and can reconstruct biomolecules in Lego-like fashion. It learns how molecules interact with each other.
FeNNix-Bio1 has proved its effectiveness in one of the most difficult tasks in molecular modelling: simulating the physical behaviour of water in its various phases. Indeed, the foundation model can accurately predict various physical properties, and reproduce the behaviour of ions and small organic molecules in solution with remarkable fidelity, where other reference models are unable to do so. This is essential because water is the solvent present in the human body, and its interaction with drugs plays a key role in their activity.
One of the special features of the model is its ability to model the reactivity of molecules, i.e. to create or break chemical bonds, which traditional simulation software cannot do. In this way, FeNNix-Bio1 makes it possible to design covalent drugs (i.e. those which bind directly to the target through the creation of a chemical bond) such as Paxlovid or Ibrutinib.
The Sorbonne University research team that developed FeNNix-Bio1 set itself the goal of going beyond the capabilities of AlphaFold, the artificial intelligence software developed by Google DeepMind, which provides a prediction of protein structure based on their amino acid sequence. But FeNNix-Bio1 goes further.
“AlphaFold has revolutionised protein structure prediction. However, proteins are not static, and their structures evolve over time, modifying drug interactions. FeNNix-Bio1 makes it possible to model these dynamic effects. In addition, AlphaFold does not accurately model the interactions of proteins with drug candidates. FeNNix-Bio1 addresses these two important limitations for biomolecular simulation” said Piquemal.
FeNNix-Bio1 is designed to offer quantum-level accuracy while remaining scalable and cost-effective. The foundation model doesn't just predict structure, it understands how molecules behave and interact.
Predicting a drug's ability to bind to a protein (or to RNA or DNA) is one of the most complex tasks in drug discovery. Chemical space is virtually limitless. An infinite number of drug molecules can be designed, and introduced into an infinite number of targets (around 100,000). There are thus trillions of possible combinations, and it's impossible to fit them all into a database.
One feature is its potential to reduce laboratory experimentation and enable the exploration of more innovative drug candidates to tackle targets that are complex or even considered impossible to modulate until now.
To achieve this, the FeNNix-Bio1 researchers developed neural network approaches adapted to applications in chemistry and physics, rather than using LLMs (Large Language Model) architectures, generally optimised for recognising and generating text. More accurate and less expensive, FeNNix-Bio1 can be trained in a few hours using a standard GPU, whereas other AI models require weeks of supercomputing time.