Artificial intelligence is beginning to change that. It isn’t a silver bullet, and no medicine yet on the market has been “discovered by AI” alone. But the technology is already reshaping how scientists work – from the questions they ask in the lab to the way clinical trials are designed. Its influence is becoming harder to miss.

That progress has taken time. As Elena Fonfria, director and project leader of biology at Recursion, puts it: “Biology is really complex and drug discovery is becoming slower and more expensive to just explore every single avenue or try to understand everything. What we do is use AI and technology to have a fundamentally new approach for the model that vastly improves speed and efficiency.”

For years, that foundation didn't exist. Today it does. And that makes all the difference.

Recursion has spent more than a decade building this foundation. The company’s phenomics and transcriptomics platforms provide complementary views into human biology: microscopy imaging that captures the broad activity across cells, and molecular data that tracks how DNA is transcribed into proteins. Together, these assays are run at an industrial scale, producing almost 20 petabytes of data.

“Phenomics gives you a view of the broad activity across a cell,” Nilsson says. “Transcriptomics clicks down into another view of the central dogma […] They’re both assays you can run at really high control and repeatability, with fit-for-machine-learning controls in order to build these models.”

She continues: “This makes things way more efficient because, rather than having to generate data every time you want a new insight, you're relying on the data you've generated previously over time. And because of the reliability of the data, the way these machine learning models work, every time you add a data point you're able to compare it against large swaths of existing data.”

The numbers are staggering. Millions of experiments can now be run under tightly controlled conditions. Petabytes of results are stored and analysed. What matters is not just volume, but repeatability. The same signal can be observed week after week, or known drug interactions show up exactly as expected. That reliability turns raw biology into something AI can truly learn from.

Computation is the other critical ingredient. Supercomputers now simulate aspects of human biology digitally, creating “world models” that don’t just describe cells, but predict how they might behave under different conditions. For Nilsson, this marks a turning point: “In this new era of foundation models and world models, that’s a big differentiator.”

Traditionally, scientists might screen half a million compounds in search of a lead. AI changes that process. Models generate predictions about which molecules could work, and those predictions are then tested in the lab. Results feed back into the system, sharpening the next set of ideas. Fonfria calls it the “predict-test-learn” cycle. It means fewer dead ends, less wasted time, and the chance to move quickly when a promising signal appears.

“In chemistry, you would need a very good chemist to explore all the chemical space – the computers can imagine the molecules automatically,” she explains.

This approach is already visible in pipeline projects. In oncology, AI has guided the design of CDK7 inhibitors and uncovered new biology around RBM39. In rare diseases, it has pointed researchers towards alternative strategies for conditions like hypophosphatasia – ideas that would have taken far longer to surface using conventional methods.

The same principle applies in clinical development. Predictive models suggest the most appropriate doses, simulate patient responses, and highlight where to find eligible participants. The effect is shorter trials and fewer burdens on patients. As Fonfria explains: “If we are sure in our predictions, if we are anchoring what we know preclinically, and our models, that means that very quickly we can give the patients doses that will be appropriate for them. Collectively, the clinical trials are shorter and better targeted for what the patient needs.”

Founding fellow at Recursion Mike Genin, who has spent more than 30 years in discovery roles primarily at large pharma companies, has seen each of these waves rise and fall.

“I’ve never seen anything live up to its initial hype, ever […] Not even genomics,” he says. “It's because it's so hard, it's so complex, and nobody ever really tried to put all the pieces of the puzzle together. There are a variety of reasons for that. I think part of the reason is we didn't have the computing power that we have now.”

So, why does he believe AI is different? Because, this time, the advances are happening together. Reliable data. High-throughput automation. Generative chemistry. Supercomputing. Instead of sitting in silos, these tools are converging.

Progress has still required patience. Nilsson remembers when some of the approaches that now deliver results were little more than cartoon diagrams on a PowerPoint slide. Early attempts failed. But coming back to the same ideas with better tools has made a difference: “There are definitely models that we tried years ago, and then we’ve come back on and had success more recently,” she recalls.

“A decade ago, most companies, including Recursion, were building models off of single datasets,” she explains. “Now, we're seeing huge improvements in performance and their ability to understand biology and chemistry by combining different datasets.

“That's because the models aren't just learning about one assay, they're learning to reason about biology and chemistry broadly. That's really new.”

That experience left him determined to pursue not incremental improvements, but transformative therapies. After decades in big pharma, he believes AI may finally deliver the step change patients need.

“I first had cancer in 1980. The field has certainly changed since then, but it has not changed enough. The drugs they gave me back in 1980, they still use today, which is good. It shows they work at times, but they're actually pretty toxic agents, as are most chemotherapeutic agents,” he says. “Rather than working on best-in-class molecules, I was always interested in first-in-class molecules, taking advantage of novel biology, novel pathways that nobody's ever done before. I thought, and believe to this day, that's where real effective additions to the current arsenal of therapies will be discovered.”

Fonfria sees the same urgency: “Patients cannot wait.” For her, the potential of AI is not just efficiency, but disruption – offering new approaches to diseases like cancer, Alzheimer’s, and Parkinson’s, where existing treatments remain limited. “The big diseases don’t have a proper cure, yet […] Getting the computers to assist us, I think it’s a brilliant approach. It’s really exciting.”

Nilsson adds that culture matters as much as code. The ability to bring biologists, chemists, engineers, and data scientists together is what unlocks breakthroughs. “It’s not a specific model, it’s not a specific dataset. It’s how you make these teams that come from different worlds and disciplines able to communicate and work closely,” she says.

For Genin, the test of success will always come back to the clinic. “Efficiency, in my mind, is speed and quality. If you’re sacrificing quality, then I’m not interested. Quality at the end of the day is what moves the needle in the clinic, and it’s what gives parents and patients hope.”

AI is no longer just a distant promise. It is already part of how scientists discover and test new medicines, and the first generation of therapies shaped by it is edging closer. Progress has taken time, as it always does in science, but the direction is clear.

While the near-term breakthroughs may not wear the badge of “AI drug” yet, their origins will be inseparable from the models and algorithms that guided their discovery. And, for patients waiting on better options, that shift cannot come soon enough.