Targeting RNA with small molecules using state-of the art methods provides highly predictive...
Scientists have long struggled with a difficult trade-off in drug design. Only very slow, costly computations have allowed one to simulate molecular interactions with quantum-level accuracy, capturing the subtle dance of electrons and atoms. Though they depend on approximations that overlook important details, faster techniques exist. Now bridging this gap, an artificial intelligence model called FeNNix-Bio1 is providing practical speed and cost-effective quantum chemical accuracy. Combining the best of both worlds in molecular simulation, this breakthrough is poised to hasten drug discovery and change how we study molecules.
But how do we break the speed versus accuracy limit?
Predicting how a protein folds or how a drug candidate will bind to its target in drug research boils down to physics and conventions have pushed researchers into a choice between speed and accuracy:
For example, explicitly treating electrons and using approaches like density functional theory (DFT) or quantum Monte Carlo produce quite accurate results. Unfortunately, they are expensive and painfully slow, so it is not feasible to run long simulations for drug screening or replicate big biomolecules.
Another example is the use of classical force fields. These are basically pre-defined equations and guidelines meant to approximatively match molecular forces. They simplify physics to allow simulations of large proteins or viruses. They run fast but at the expense of precision.
Furthermore, standard force fields often ignore quantum effects and chemical reactions, like bonds forming or breaking, and require tedious human fine-tuning of parameters.
FeNNix-Bio1 changes all of this by learning the complex laws of physics directly from data. It uses a neural network, which is an AI model trained on a database of quantum chemistry calculations, and in effect acts like a quantum calculator. FeNNix-Bio1 uses machine learning to predict the energy and forces on atoms, which allows it to achieve high accuracy at speeds rivalling traditional simulations.
The model was trained on synthetic quantum data by using theory instead of experimental measurements. This in effect created the “perfect” teacher for AI to understand molecular interactions without human error.
It’s akin to having AI study a grandmaster playing chess before embarking on its own matches. FeNNix-Bio1 learns the moves, in this case the laws of physics, and can then predict molecular behaviour in whatever scenario. This approach isn’t limited to one type of molecule as it’s a foundation model. This means it can be widely applied to many chemical systems and improve itself as more data becomes available.
FeNNix-Bio1 vs. Traditional Methods: A Quantum Leap
What advantages does FeNNix-Bio1 bring over established methods? First, it sidesteps the laborious parameter tuning of classical force fields. Normally, experts spend years tweaking parameters so that simulations match experimental data for, say, water or proteins. FeNNix-Bio1 doesn’t need such manual adjustment. It learns directly from physics, which greatly reduces human effort.
Second, unlike classical models, it naturally handles chemical reactivity. This means if a simulation involves a bond breaking, like a drug molecule reacting, FeNNix-Bio1 can capture that spontaneously. Previously, this required costly quantum chemistry on-the-fly or specialized reactive force fields.
Despite its strong physics grounding, FeNNix-Bio1 runs blazingly fast. The researchers report that its inference speed, or how quickly it computes forces, is on par with optimized advanced force fields used in today’s molecular dynamics simulations. In one benchmark, FeNNix-Bio1 achieved 10× to 100× faster performance than a recent state-of-the-art models under the same conditions.
In practical terms, that increased speed lets scientists simulate larger systems or longer timescales without sacrificing accuracy. FeNNix-Bio1 can scale up to simulate hundreds of thousands or even millions of atoms on modern GPUs, tackling systems as big as an entire virus’s protein shell or a cellular membrane patch. These were sizes once confined to coarse methods.
Crucially, this AI isn’t just faster; it’s extremely accurate. In tests, FeNNix-Bio1 reached the precision of the community’s most trusted physics-based models for molecular interactions, even outperforming some of them on key metrics, all at a fraction of the computational cost. In other words, it delivers quantum quality results without the associated wait.
This combination of fidelity and efficiency is what makes experts describe FeNNix-Bio1 as a potential “game-changer” for computational chemistry and drug design.
From Water to Proteins: Proving Its Mettle
The developers put the model through a battery of tests that span the breadth of biochemistry, from the simplest solvent to complex proteins and reactions, to show that it's not a one-trick pony.
Here’s an overview of its capabilities:
- The properties of water: This solvent is infamously difficult to simulate. Most advanced techniques have struggled to precisely determine its properties, like density or boiling point temperature. FeNNix-Bio1 perfectly captured a range of water’s characteristics across different temperatures. For example, it matched experimental values for density and the energy needed to evaporate water (the so-called enthalpy of vaporization) . It also replicated the liquid water’s detailed structure and how water molecules arrange themselves around one another. Furthermore, the model predicted water’s hydration free energy as –6.49 kcal/mol. This is virtually the same as the experimental measure of –6.32 kcal/mol . This accuracy is rare and shows that the AI is capturing subtle effects rather than ignoring them.
- Ions in Water and small organic molecule free energies of hydration: Salt ions, like sodium and chloride, are common in biology, so having accurate simulations of their interactions with water is essential. When tested on ions in solution, the model replicated the subtle behaviour that these charged particles displayed. For example, it correctly captured how ions organize around the surrounding water molecules. In fact, the ion-water distances it produced closely matched those obtained using the best classical force fields, except those were painstakingly tuned for ions. The model can handle charges and many-body effects without tweaks. Moreover, FeNNix-Bio1 computed the hydration free energies of ions in line with experimental values. This is yet another benchmark showing both accuracy and transferability of the model from one environment to another.
Such benchmark has been extended to the evaluation of small organic molecule free energies of hydration. Hydration free energies (HFEs) represent the free energy difference between a molecule solvated in water and the same molecule in the gas phase. It is a key quantity in drug discovery as it is a building block for computing binding affinities. Here also, FeNNix-Bio1 provides excellent predictions within the range of the chemical accuracy compared to experiments.
- Protein Folding Dynamics: The models must control large molecules and their movements, so the team used FeNNix-Bio1 to fold and unfold a small protein. The AI could capture the protein’s folding free-energy landscape, which maps out the stable and unstable shapes the protein can take. This is significant because protein folding is a delicate balance of forces and many intermediate steps. Having a simulation with quantum-level accuracy gives researchers a greater understanding how structure emerges from sequence. The folding data obtained by FeNNix-Bio1 aligned with known results, suggesting the model can scale towards more complex and flexible macromolecules.
- Protein–Ligand Binding: Drug design often hinges on binding free energy, or drug docking. Highly accurate calculations usually require complex methods and a substantial amount of computer time. FeNNix-Bio1 determined a small molecule’s absolute binding free energy when binding to a protein. The result was –5.5 kcal/mol, which is within a hair’s breadth of the experimental value of –5.44 kcal/mol. A difference of around 0.1 kcal/mol is within experimental error. This means the AI’s prediction was as good as a lab measurement. This level of precision in binding studies means drug designers could rely on FeNNix-Bio1 to virtually screen and rank drug candidates with confidence and remove long-shot compounds before any lab testing.
- Impromptu chemical reactions: FeNNix-Bio1 has the ability to model chemical reactions, which standard biomolecular simulations can’t normally handle. Quantum chemistry can, but it’s impractical beyond very small systems. The researchers challenged FeNNix-Bio1 with a classic reaction: the conversion of cyclobutene to butadiene (a ring-opening reaction). The AI model captured the reaction’s pathway and estimated the energy barrier to give a semi-quantitative picture of the reaction without any special training on reaction data. Essentially, the model understood chemistry enough to handle a reaction and opens the door to studying enzymatic reactions or drug metabolism in silico.
Faster and more intelligent drug discovery:
By combining real-world speed with quantum accuracy, this AI model has the potential to drastically reduce the time needed to develop new drugs:
In drug discovery’s early phases, it’s commonplace to sift through thousands of possible compounds to find those that bind to a target protein efficiently. FeNNix-Bio1 allows researchers to perform much more accurate screening.
The AI can model the atomic-level interactions between a protein and a drug candidate, including minute quantum mechanical effects. This allows researchers to assess many possible compounds and narrows the list of candidates before synthesis and testing. This virtual lab focuses resources on only the most promising ones but at a fraction of the cost and time of real experiments.
- Less reliance on experimentation: Drug design still relies heavily on trial and error. However, FeNNix-Bio1 can predict failures or adverse reactions early, which vastly reduces the dead ends in later stage drug development. This could save drug developers tens of millions of dollars on compounds destined for failure.
- Synergies with tools: FeNNix-Bio1 arrives at a time when other breakthroughs, like generative AI, are changing the world. The model acts as a missing link to evaluate and refine those proposals created with generative AI in physics-based detail. Scientists can take an AlphaFold-predicted protein structure and immediately start a simulation with FeNNix-Bio1 to see how that protein might move and bind. This streamlines the process for bringing structural insights for molecules to delivering promising drugs candidates.
- Towards Quantum (Computing) advantage: The AI model also has synergies with quantum computing. While the tech is still evolving, FeNNix-Bio1 will benefit from the emergence of computers that can accelerate and enhance quantum chemistry computations. The team anticipates using quantum computing to generate richer datasets, especially those for complex molecular interactions. This would nudge the model towards absolute chemical accuracy.
- A Foundation for Broad Scientific Exploration: While the spotlight is on drug design, the applications are far wider as FeNNix-Bio1 is a foundation model for molecular simulations. This means it provides a general platform that can be adapted to countless scenarios. For example, chemists could use it for material science when designing better battery electrolytes. The FeNNix-Bio1 team envisions using quantum computing power to further enhance the training data.
- Speed with validated accuracy: The AI model isn’t just fast. When measured against hydration free energy benchmarks, its accuracy is on par with experimental results. FeNNix-Bio1 precisely captures how ions behave in water without the need to manually fine-tune the model. This allows it to scale effortlessly and simulate molecular systems for much larger structures.
FeNNix-Bio1 lowers barriers to entry. Universities are unlikely to have the expertise or manpower to develop custom force fields or run huge quantum computations. However, this AI model can be accessed by researchers anywhere to run high-fidelity simulations.
The Future of In Silico Science
FeNNix-Bio1's development introduces an era where physics and AI can collaborate to solve challenges that were once thought impossible. We are in foothills of what could become a standard tool in laboratories worldwide: AI models acting like computational microscopes and clearly showing accurate molecular behaviour. This could result in the faster development of therapies or even personalized medicine by fast simulating how different drug candidates behave in different protein variants. There will be fewer unexpected pitfalls in clinical trials because the molecular surprises in drug discovery would be greatly mitigated. Like in engineering, science as a whole emphasises a paradigm shift whereby in silico studies guide and augment real-world experiments more tightly than ever.
FeNNix-Bio1’s success also shows the force of multidisciplinary creativity. This model blends expertise and advances across domains as varied as quantum chemistry, machine learning, and high-performance computing.
The outcome is way greater than the sum of its components. A tool that has learned the rules of physics and quantum chemistry can now run "what-if" scenarios at will. It’s like a chemical flight simulator that flawlessly knows and applies the laws of physics, rather than aerodynamics, to empower researchers to test their ideas in a safe, quick and affordable digital environment without committing to the time, costs and safety risks associated with real-world trials.
In drug discovery, where time literally equates to life, having such flexibility is crucial. FeNNix-Bio1 brings us a giant leap closer to the long-held dream of truly rational drug design, where we can design molecules with full knowledge of how they will behave in the complex environment of the human body.
From enabling new science discoveries to speeding up the delivery of life-saving medications, this model is an example of how an AI discovery might spread outward. FeNNix-Bio1 is still in development and as it improves the scientific community will continue observing it closely. A revolution in how we replicate the molecular world in a digital platform could inspire ideas we have only dreamed about. It could usher in a future where discovery is constrained less by what we can do and more by what we can muster with human creativity.
Access the full paper here
