Targeting RNA with small molecules using state-of the art methods provides highly predictive...
AI and Quantum Chemistry: Powering the Next Leap in Drug Design
Developing effective treatments starts with understanding how drug-like molecules interact with biological targets. This is driven by molecular dynamics—the precise choreography of atoms forming and breaking bonds in real time. For decades, simulating this was slow, costly, and complex. But with advances in AI and physics-based modelling, this once-impossible dance is now faster, smarter, and more accessible—transforming how we design drugs, materials, and beyond.
In April 2025, Sorbonne University and Qubit Pharmaceuticals researchers revealed FeNNix-Bio1, an AI foundation model that enables next-generation molecular simulations crucial to drug discovery. The model has been developed with the supercomputing support of Argonne National Labs, EuroHPC and GENCI. This transformational advancement provides a new approach for molecular science at scale.
Traditional molecular simulations depend on empirical force fields. These sophisticated formulas require extensive fine-tuning by computational scientists. While already powerful, they tend to be limited in accuracy and scope. At the other end of the spectrum, quantum chemistry models are highly accurate but far too slow and expensive to implement at scale. The holy grail is to have it all: rapid, scalable prediction with quantum-level accuracy. That's what FeNNix-Bio1 aims to achieve.
“FeNNIx-Bio1 extends the breakthrough in protein structure prediction introduced by AlphaFold to any biomolecular simulation," says Jean-Philip Piquemal, Qubit’s Chief Scientific Officer and the Distinguished Professor behind the FeNNix-Bio1 team. AlphaFold revolutionized protein structure prediction; however, it doesn't capture how small drug-like molecules interact with those proteins—this is where the physics still matters.
Accurately modelling drug–protein interactions isn’t just about smarter algorithms—it’s about navigating chemical space at scale. The space of potential combinations is unimaginably vast. Scientists can design countless drug-like molecules, each with the potential to bind to different proteins in unique ways. There are trillions of possible combinations. You can't fit them all into a database. “It’s like trying to fit the universe into a spreadsheet,” says Piquemal.
So how do you deal with the trillions of potential permutations? You compute. While most earlier models were built on decades of accumulated experimental data—like X-ray crystallography and NMR spectroscopy—FeNNix-Bio1 is trained on synthetic data generated from quantum chemistry simulations rooted in first principles. This means, it relies on the fundamental laws of physics—like Schrödinger’s equation—to simulate how electrons and atoms behave in molecules. This allows the model to learn directly from the underlying physics rather than from biased, noisy or incomplete experimental datasets.
“This unbiased approach does not focus on what has been previously discovered. Instead, it learns how molecules should behave, based on the underlying physics and chemistry laws," Marino adds. “At the same time, it's flexible—evolve the building blocks, and you can model just about anything.”
Instead of generic LLMs, the FeNNIX-Bio1 team developed custom, innovative AI frameworks for chemistry and physics based on neural networks. "LLMs perform well with language and have not been designed to learn chemistry or physics," says Piquemal. "They're also very expensive to train. Our model is cost-effective and more focused."
FeNNix-Bio1 can be trained in just a few days using a single GPU card. Other AI models in this class require weeks of supercomputer time. With this more accessible training environment, the model compares to, and is often superior to the best force fields used today in pharma while being extremely scalable thanks to GPU-accelerated inference.
"This is a plug-and-play substitute for current methods," says Robert Marino, Qubit's CEO. "Whereas earlier models required five to ten years to build and implement, FeNNix-Bio1 can be trained and implemented in days with less human intervention."
To test how well FeNNix-Bio1 would do, the researchers challenged it with one of molecular science's most deceptively difficult tasks: modelling water. The model excelled, accurately predicting water physical properties as well as how small molecules interact with it outperforming conventional methods while closely matching real-world experimental data.
The model also predicted chemical reactions, such as the rearrangement of butadiene to cyclobutene, something traditional force field simulations cannot achieve, showing that FeNNix-Bio1 understands molecular behavior and changes.
The most exciting feature is its potential to reduce physical experimentation and allow for the exploration of more innovative biomolecules to address challenging or undruggable targets. “As the model is as accurate as the experiment, we can fail quickly and cost-efficiently in silico before going to the lab,” says Piquemal.
Additionally, FeNNix-Bio1 improves with new synthetic data with minimal overhead and is compatible with other structure prediction foundation models enabling an AI-driven end-to-end design cycle: predict protein structure, design drug candidates, and model interactions. Its integration with Alphafold-like models (MIT’s Boltz-1 approach was used in the study) allows researchers to shift from predicted protein structures to accurate molecular simulations that exactly mimic how drugs and proteins interact.
There’s also a deeper tech story here. FeNNix-Bio1 opens the door to quantum AI, which is the convergence of quantum computing, high performance computing (HPC) and machine learning. FeNNix-Bio1 sets new benchmarks for speed while simulating environments with as many as several million atoms. It can do this without lowering the quantum accuracy. While quantum computers are still maturing, we can still revolutionize data generation for molecular simulations with quantum chemistry combined to HPC and “we’re already starting to use quantum-computing algorithms to generate data to enhance our models,” says Piquemal. Leveraging partnerships with quantum hardware companies, more data will continuously be added in the future.. “What people thought would take until 2035 is already starting to happen today.”
FeNNix-Bio1 is changing the way we design, test, and understand molecules. It’s a strategic advantage, with FeNNix-Bio1, we can move from scientific curiosity with respect to quantum chemistry to a practical implementation with AI. First, we tackle a more cost-effective approach for the molecular dance within drug discovery with subsequent use to industrial enzymes, materials science, desalination membranes, and even battery design.
FeNNix-Bio1 is a paradigm shift in designing, testing, and understanding molecules. It's not only quicker, it's also smarter, more efficient, and ready to scale to industries. With AI-enabled biology at scale and quantum computing shattering physics boundaries, the long-awaited vision of automated discovery of molecules is becoming a reality at an accelerating rate.
Access the full paper here