Gene expression provides an additional layer of information that complements traditional risk factors.
For generations, medicine has struggled to explain why two people with identical risk profiles can face entirely different fates when it comes to blood clots. Researchers in Barcelona have now built an artificial intelligence system that reads the hidden molecular language of a person's genes — identifying 494 genetic signals, many of them barely studied, that distinguish those who will develop venous thrombosis from those who will not. The work does not yet belong in the clinic, but it marks a meaningful step toward a medicine that listens to the body's own story rather than relying solely on the blunt instruments of age, weight, and family history.
- Venous thrombosis kills and disables thousands each year, yet in the majority of cases the known risk factors fail to explain who will actually clot — leaving doctors to guess and patients to slip through the cracks.
- A Barcelona research team built an AI that layers clinical data, genetic variants, and gene expression profiles together, teaching itself to recognize patterns of risk that no single data stream could reveal alone.
- The system uncovered 494 genes — including long noncoding RNAs, regulatory molecules almost entirely unstudied in clotting disease — as meaningful signals separating thrombosis patients from healthy individuals.
- Adding gene expression data cut false high-risk classifications nearly in half, from 43% to 23%, while simultaneously improving detection of true thrombosis cases from 70% to 74%.
- The tool is not yet validated for clinical use, but it opens a path toward personalized prevention — identifying who genuinely needs intervention and sparing others from unnecessary treatment and anxiety.
Two patients walk into a clinic with identical charts — same age, same family history, same weight — yet one will develop a life-threatening blood clot and the other never will. Medicine has long lacked a satisfying answer for why.
Researchers at the Sant Pau Research Institute in Barcelona set out to change that. They built an AI system capable of reading not just the familiar risk factors, but the deeper molecular language encoded in a person's genes. Published this summer in the Journal of Thrombosis and Haemostasis, the work identified 494 genes whose activity patterns predict who is likely to develop venous thrombosis — a condition responsible for thousands of deaths and serious injuries each year.
Venous thrombosis has known triggers: obesity, certain hormones, age, specific genetic mutations. Yet these explain only part of the picture. More than 60 percent of cases have a genetic dimension, but the known hereditary markers still leave many unexplained — particularly so-called idiopathic cases, where no obvious cause exists at all. The team analyzed 790 people from families with clotting histories, feeding the AI three layers of data: clinical measurements, genetic variants, and gene expression profiles showing which genes were active or suppressed in each person's cells.
The results were striking. Among the 494 genes identified, many were long noncoding RNAs — regulatory molecules that have barely been studied in the context of clotting. When the AI incorporated gene expression data alongside clinical and genetic information, false high-risk classifications fell from 43% to 23%, while accurate detection of true thrombosis cases rose from 70% to 74%. In medicine, that combination — fewer false alarms, more real catches — is precisely what better tools are meant to deliver.
The researchers distilled their findings into a molecular signature: a composite score measuring how closely a person's biological profile resembles that of someone who has already clotted. This allowed them to flag individuals with no clotting history whose molecular patterns quietly resembled those of patients who had. The data also surfaced connections to the heart and kidneys, reinforcing links that earlier research had only hinted at.
Director Dr. José Manuel Soria was clear that the tool needs validation in independent patient populations before it can enter clinical practice. But the direction is visible: a future where doctors read the molecular story written in a patient's cells, identifying who truly needs prevention and who can be spared unnecessary worry — and where the newly discovered genes become not just predictors, but targets for understanding how thrombosis begins and how it might one day be stopped.
Two people walk into a clinic with identical medical charts. Same age. Same family history. Same weight. Yet one will develop a blood clot that could kill them, and the other will live their whole life without incident. Doctors have never had a good answer for why.
Researchers at the Sant Pau Research Institute in Barcelona decided to stop guessing. They built an artificial intelligence system that reads not just the obvious risk factors—the ones we've known about for decades—but the hidden molecular language written in a person's genes. The work, published this summer in the Journal of Thrombosis and Haemostasis, identified 494 genes whose activity patterns can predict who will develop venous thrombosis, a condition that kills or permanently injures thousands of people every year.
Venous thrombosis—blood clots in the veins—ranks among the most common cardiovascular killers. We know some triggers: obesity, certain hormones, age, specific genetic mutations. But these factors explain only part of the story. In more than 60 percent of cases, genetics plays a role, yet the known hereditary markers don't account for why some people clot and others don't. The most puzzling cases are those with no obvious trigger at all, called idiopathic venous thromboembolism. These are the patients who slip through the cracks of conventional screening.
The team analyzed 790 people from families with a history of clotting disease, including 70 who had experienced unexplained thrombosis. They fed the AI system three layers of information: clinical data like age and body mass index, genetic variants, and gene expression profiles—essentially, which genes were turned up or down in each person's cells. The machine learned to recognize patterns invisible to traditional analysis. It found that von Willebrand factor levels, BMI, age, and certain ABO blood group variants remained important predictors. But it also discovered that 494 genes—many of them long noncoding RNAs, regulatory molecules barely studied in the context of clotting—carried crucial information about risk.
The improvement was striking. When the AI used only clinical and genetic data, it wrongly classified 43 percent of healthy people as high-risk. Add gene expression information, and that false alarm rate dropped to 23 percent. At the same time, the system's ability to correctly identify people who had actually experienced thrombosis improved from 70 percent to 74 percent. That may sound modest, but in medicine, reducing unnecessary treatment while catching more real cases is the whole game.
The researchers created what they call a molecular signature—a composite score that measures how closely a person's biological profile resembles someone who has already had a clot. This allowed them to spot people with no history of thrombosis whose molecular patterns looked suspiciously like those of patients who had clotted. The work also reinforced connections between thrombosis and other organ systems: the heart and kidneys both showed up in the data, confirming what previous research had hinted at.
Dr. José Manuel Soria, who directed the work, emphasized that the tool is not yet ready for the clinic. It needs testing in independent groups of patients to prove it works beyond this single study population. But he and his colleagues see the direction clearly: a future where doctors don't just count risk factors on a checklist, but read the molecular story written in a patient's cells. That story, they argue, will let us identify who truly needs prevention and who can be spared unnecessary worry and medication. The genes and regulatory molecules they found represent not just better predictions, but new targets for understanding how thrombosis actually develops—and eventually, how to stop it.
Notable Quotes
The study's main contribution is not only the identification of new genes associated with thrombosis, but also the demonstration that integrating thousands of biological variables makes it possible to describe risk profiles far more accurately than when traditional factors are analyzed in isolation.— Dr. José Manuel Soria, director of the Complex Disease Genomics Unit at IR Sant Pau
The incorporation of transcriptomic data allowed us to identify disease-associated signals that could not be detected using conventional approaches. This demonstrates the potential of combining artificial intelligence and gene expression to achieve a more precise characterization of patients.— Dr. Pol Ezquerra, first author of the study
The Hearth Conversation Another angle on the story
Why does the AI do better than doctors just looking at the standard risk factors?
Because it's reading thousands of biological signals at once instead of a handful. A doctor sees age, weight, family history. The AI sees which genes are active or quiet in your cells—it's like the difference between knowing someone's address and actually walking through their house.
But aren't genes fixed? Why would gene expression matter if you're born with the same DNA as everyone else?
Your DNA is fixed, but which genes are switched on or off changes based on your health, your environment, what's happening in your body right now. Two people with identical genes can have completely different gene expression patterns. That's where the real variation lives.
The numbers seem modest—from 70 to 74 percent detection. Is that actually meaningful?
In thrombosis, yes. You're talking about catching four more people per hundred who will actually clot. Those four people might get preventive treatment that saves their lives. Meanwhile, you're also cutting false alarms in half, so you're not treating people who don't need it.
What about those long noncoding RNAs they mention? Why haven't we studied those before?
They're harder to measure and they don't code for proteins, so they were invisible to older research methods. But they regulate other genes—they're like the instructions for the instructions. The AI can spot their patterns even when we don't fully understand what they do.
So when will patients actually benefit from this?
Not yet. It needs to be tested on completely different groups of people first. But if it holds up, maybe in five years you could get a blood test that reads your gene expression and tells you your real risk. That's the promise.