Jump between scenes in a molecular movie, not watch every frame
In laboratories at Chalmers University of Technology and the University of Gothenburg, Swedish researchers have built an artificial intelligence that learns the grammar of molecular motion rather than computing its every syllable — moving ten thousand times faster than the methods that have long governed drug discovery. The work touches something enduring in the human effort to heal: the gap between what we can imagine and how long it takes to find out if we are right. By teaching a machine to recognize the deep patterns beneath molecular change, these scientists have opened a path toward compressing years of early-stage pharmaceutical searching into something far closer to the speed of curiosity.
- Drug development's most expensive bottleneck — screening thousands of molecular candidates one painstaking atomic step at a time — has long consumed years and enormous resources before a single promising compound emerges.
- Traditional molecular dynamics simulations move in femtosecond increments, forcing researchers to run billions of calculations just to observe biological processes that unfold over far longer timescales.
- The TITO model sidesteps this exhaustion entirely, learning the underlying rules of molecular motion so it can leap across timescales a thousand times longer than its training data — without recalculating every atomic interaction.
- Validated against over 12,500 organic molecules and peptides, the model's predictions held up against classical simulations, satisfying the physics consistency that real laboratory guidance demands.
- Pharmaceutical companies are already paying close attention, sensing that a tool capable of compressing months of computation into seconds could reshape how quickly new medicines reach patients.
Swedish researchers at Chalmers University of Technology and the University of Gothenburg have developed an AI model capable of predicting molecular behavior more than ten thousand times faster than conventional methods — a development that could transform the earliest and most costly phase of drug discovery.
Bringing a new medicine to patients typically spans more than a decade, with much of that burden concentrated at the screening stage, where vast numbers of molecular candidates must be evaluated before a handful worth pursuing can be identified. The standard tool for this work, molecular dynamics simulation, advances in femtosecond increments — a quadrillionth of a second at a time — demanding billions of steps to capture processes that matter for drug development. The computational cost is immense.
The new model, called TITO, takes a fundamentally different approach. Rather than recalculating every atomic movement, it learns the governing principles behind molecular change. Research leader Simon Olsson describes it as jumping between key scenes in a molecular film rather than watching every frame. Crucially, the model generalizes — it can make predictions about molecules it has never encountered, because it has internalized rules rather than memorized specific cases.
Tested on more than 12,500 organic molecules and over a thousand short peptides, TITO predicted molecular behavior across timescales a thousand times longer than its training data, with results that held up against traditional simulations and remained consistent with physical law. Lead author Juan Viguera Diez, an industrial doctoral student at AstraZeneca, underscores that the breakthrough lies in demonstrating AI can capture molecular physics in a genuinely general way.
The model currently operates on relatively simple systems at a single temperature, and the team is working to extend it toward more complex, realistic conditions. But the trajectory is clear: a tool that condenses months of computational labor into seconds carries the potential to fundamentally accelerate how humanity searches for its next medicines.
Swedish researchers have built an artificial intelligence model that can predict how molecules behave over time at a speed that is more than ten thousand times faster than the traditional methods chemists have relied on for decades. The work, published in Science Advances by teams at Chalmers University of Technology and the University of Gothenburg, could reshape the earliest and most expensive phase of drug development—the screening stage where thousands of molecular candidates are tested to find the few worth pursuing further.
Bringing a new drug to patients typically takes more than a decade, and much of that time and cost is concentrated at the beginning. Researchers must test vast numbers of molecules to identify which ones show promise. The conventional approach, called molecular dynamics, requires scientists to calculate the forces between every atom in a molecule and then move them forward in tiny increments—each step lasting about one femtosecond, or a quadrillionth of a second. Because the biological processes that matter for drug development unfold over much longer periods, simulations demand billions of these microscopic steps, making the work computationally exhausting and slow.
The new model, called TITO (Transferable Implicit Transfer Operators), works differently. Instead of calculating every atomic movement, it learns the underlying rules that govern how molecules change. The researchers trained it on simulated examples of molecular motion, allowing it to recognize patterns in how atoms move and rearrange. Once trained, the model can predict how new molecules will behave without having to compute each step. Simon Olsson, the research leader and associate professor at Chalmers, describes it as jumping between scenes in a molecular movie rather than watching every frame in sequence. The model learns general principles rather than memorizing specific systems, which means it can make predictions about molecules it has never encountered before.
To test the approach, the team examined more than twelve thousand five hundred organic molecules—compounds built from carbon, nitrogen, hydrogen, and oxygen—along with over a thousand short peptides, the building blocks of proteins. The model proved remarkably capable: it could predict molecular behavior over timescales a thousand times longer than the data it had seen during training. When researchers validated these predictions against traditional simulations, the results aligned. The model's forecasts remained consistent with the laws of physics, a crucial requirement for any tool meant to guide real laboratory work.
Juan Viguera Diez, an industrial doctoral student at AstraZeneca and the lead author of the study, emphasizes that the breakthrough lies in demonstrating that artificial intelligence can capture the underlying physics of molecular systems in a general way. The pharmaceutical industry has taken notice. Companies are eager for simulations that more accurately reflect reality and allow them to identify promising drug candidates faster. The researchers believe this model represents a significant step toward more efficient early-stage screening, where the ability to test larger numbers of molecules more quickly could accelerate the entire development timeline.
The work is still in its early stages. The model has been tested on relatively simple molecular systems in simplified solvent environments at a single temperature. The team is now working to extend it to more complex and realistic conditions. But the potential is clear: a tool that can compress months of computational work into seconds could fundamentally change how chemists search for the next generation of medicines.
Citas Notables
The model learns the underlying dynamics over longer time scales. It not only provides insights into the shapes that molecules take on, but also into how quickly and through which pathways these molecular transitions occur.— Simon Olsson, research leader and Associate Professor at Chalmers University of Technology
In the long term, AI models like ours could help to identify promising drug candidates more quickly and improve accuracy in the early stages.— Juan Viguera Diez, lead author and industrial doctoral student at AstraZeneca
La Conversación del Hearth Otra perspectiva de la historia
Why does it matter that this model is ten thousand times faster? Isn't speed just a convenience?
It's not just convenience—it's about access. Right now, researchers can only test a fraction of the molecules they'd like to because the computation is so expensive. Speed means they can screen far more candidates, which means better odds of finding something that actually works.
But the model is trained on simulations, not real molecules. How do we know it's predicting anything true?
That's the careful part. They validated it against traditional simulations—the slow, expensive kind that we know are accurate. The predictions matched. And the model respects the laws of physics, which is non-negotiable. It's not guessing; it's learned the rules.
The article mentions it can predict behavior over timescales a thousand times longer than what it trained on. That seems almost too good to be true.
It does sound magical, but it's because the model isn't memorizing individual molecules. It's learned the underlying patterns of how atoms move and interact. Once you understand the grammar, you can read sentences you've never seen before.
What's the catch? Why isn't this already being used in every pharmaceutical lab?
It's only been tested on relatively simple systems so far—small molecules in simplified conditions. Real drug development involves messy, complex environments. The researchers are working on scaling it up, but that's still ahead of them.
So this is a proof of concept that might change everything, but not quite yet.
Exactly. It shows what's possible. The pharmaceutical companies are watching closely because they understand what this could mean for their timelines and their costs.