We designed new molecules using AI based on the effect we wanted to achieve
For generations, medicine has hunted disease by naming its molecular villain first — but some illnesses refuse to offer one. At the Structural Bioinformatics and Network Biology Lab in Barcelona, Dr. Patrick Aloy's team has turned the question around, asking not 'what is broken?' but 'what outcome do we want?' — and training artificial intelligence to design molecules that achieve it. Their work with pancreatic cancer cells suggests that beginning with the desired effect, rather than a known target, may open paths that conventional drug discovery cannot find.
- Many cancers — including pancreatic cancer — have no single clear molecular target, leaving traditional drug discovery without a foothold.
- Aloy's team built a cellular map from over 11,000 compounds tested across cancer and healthy cell models, then trained AI to read that map and predict drug behavior.
- A generative AI platform was tasked with inventing molecules that kill cancer cells while sparing healthy ones — and the lab confirmed the designs actually worked.
- The AI-designed compounds outperformed conventional screening and carried chemical structures unlike anything in existing pharmaceutical literature.
- The methodology is still in early stages — validated in cell cultures, not yet in animals or humans — but the door it opens is significant.
Drug discovery has long followed a familiar logic: identify the broken protein, design a molecule to fix it, and test until something holds. It is a method that has served medicine well — but it depends on a disease having a clear molecular culprit. Many do not.
Dr. Patrick Aloy and his team in Barcelona chose to invert the problem. Rather than starting with a target and searching for a drug, they asked what effect they wanted — a cancer cell dying — and worked backward to find the molecule that could cause it. This approach, known as phenotypic discovery, is not new in concept. What is new is that Aloy's group used artificial intelligence to design entirely original compounds this way, then proved in the laboratory that they work.
The foundation was a database the team built themselves: more than 11,000 chemical compounds, each tested across six pancreatic cancer cell models and two normal control cell lines. That painstaking work produced a behavioral map of how molecules act in different cellular contexts. Machine learning models trained on this map proved far more predictive than older structure-based comparison methods.
With the system trained, the researchers connected it to a generative AI platform and gave it a precise brief: invent molecules active against pancreatic cancer cells but largely harmless to normal ones. The AI proposed candidates. The lab tested them. The results were striking — the designed molecules outperformed conventionally screened compounds and displayed chemical structures genuinely unlike known drugs.
Aloy is careful to frame this as early-stage work. These are candidates, not medicines, and cell cultures are not human bodies. But the methodology points toward something consequential: for diseases where the molecular landscape is too complex or too murky for target-first approaches, an effect-first paradigm guided by AI may find what traditional methods cannot.
For decades, drug discovery has followed a well-worn path: find the broken protein, design a molecule to fix it, test it in cells and animals, and hope it works in humans. It is a logical approach, and it has produced many of the medicines we rely on. But it assumes something that is not always true—that the disease has a clear molecular culprit waiting to be identified and targeted.
Dr. Patrick Aloy and his team at the Structural Bioinformatics and Network Biology Lab in Barcelona decided to flip the problem around. Instead of starting with a protein target and working backward to find a drug, they asked: what if we started with the effect we want—a cancer cell dying, or a tumor cell stopping its growth—and worked backward to find the molecule that causes it? The approach has a name in the field: phenotypic discovery. And for the first time, Aloy's group has used artificial intelligence to design entirely new chemical compounds this way, then proved in the laboratory that they actually work.
The team began by building their own database. They took more than 11,000 chemical compounds and tested each one across eight different cell models—six derived from pancreatic cancer patients and two normal control cells. This was laborious work, but it gave them something invaluable: a map of how each molecule behaved in each cellular context. They fed this information into machine learning models that could predict, with surprising accuracy, how a new compound would act in a given cell type. These predictive models proved far more reliable than older methods that simply compared the chemical structure of one compound to another.
Once the system was trained, the researchers integrated it into a generative AI platform—the kind of system that can propose entirely new molecular designs. They set it a specific task: invent molecules that are active against pancreatic cancer cells but largely inactive against normal cells. The AI generated candidates. The researchers then took many of these designs into the wet laboratory and tested them experimentally.
What they found was striking. The AI-designed molecules not only worked as intended—they showed stronger activity against the target cancer cells than compounds discovered through conventional screening methods. More than that, many of them had chemical structures unlike anything in the existing pharmaceutical literature. They were genuinely novel, not minor variations on known drugs.
This matters because many diseases, including many cancers, do not have a single obvious molecular target. The disease emerges from complex interactions across multiple proteins and cellular pathways. Traditional drug discovery can stall in these contexts, unable to identify which protein to attack. Phenotypic discovery sidesteps that problem. It says: we do not need to know the mechanism. We just need to know the effect we want, and we can design molecules to achieve it.
Aloy emphasizes that this is still early work. The compounds are candidates, not yet medicines. They have been tested in cell cultures, not in animals or humans. But the methodology opens a door. It suggests that for diseases where the molecular landscape is murky or multifaceted, artificial intelligence can accelerate the search for active compounds—and do so in a way that produces molecules more selective and more structurally innovative than traditional screening would find. The paradigm is shifting from target-first to effect-first, and the laboratory results suggest the shift is worth pursuing.
Notable Quotes
For the first time, we have designed new chemical entities using artificial intelligence based on the biological effect we wanted to achieve, and we have experimentally demonstrated that they work on specific cells.— Dr. Patrick Aloy, IRB Barcelona
The Hearth Conversation Another angle on the story
Why does it matter that these molecules are structurally novel? Couldn't an old drug work just as well?
An old drug might work, but it probably wouldn't work as well. If a molecule has never been made before, it means it occupies chemical space that conventional screening never explored. That's where selectivity often hides—in the places nobody thought to look.
So you're saying the AI found molecules in places humans wouldn't have searched?
Exactly. Human chemists design molecules based on intuition and precedent. They build on what they know works. The AI has no such bias. It can propose structures that seem odd or unlikely but that the data says will do the job.
But how do you know the AI isn't just getting lucky? That it won't fail when you test it in real tumors?
You don't, not yet. That's why this is early stage. But the fact that multiple AI-designed molecules matched their intended function in the lab—that they were selective, that they outperformed conventional compounds—suggests the system is learning something real about the relationship between molecular structure and cellular behavior.
What happens to the compounds now?
They move into further testing. Animal models, probably. Then, if they show promise, the long road toward clinical trials. But the real win here is the methodology. If it holds up, it changes how we approach diseases where we don't have a clear target to aim at.
And there are a lot of those?
Many. Most cancers, for one. Most neurodegenerative diseases. Diseases that emerge from broken networks, not broken proteins. This approach is built for exactly those problems.