PATRICK SCHWAB is no ordinary pharmaceutical researcher, and his workplace is no ordinary pharmaceutical laboratory. There are no benches or bubbling liquids. There's no white coat either. Instead, Dr. Schwab wears all black. But it's appropriate attire for someone whose workplace is in King's Cross. This area was once home to railway yards and industrial buildings, but has now been transformed into one of London's trendiest areas.
Dr. Schwab works for GSK. Pharmaceutical company. His job is to reimagine the future of drug manufacturing using an equally trendy field of computer science: artificial intelligence (AI). He applies this to drug design in silico, rather than in vitro, by shifting as much of the load as possible from the glassware to the computer.
To this end, he is developing a software tool called Phenformer and training it to read genomes. By linking genomic information with phenotypes (a biological term that describes the physical and behavioral outcomes of specific gene combinations), Phenformer learns how genes cause disease. This allows the generation of new hypotheses about the disease and its underlying mechanisms.
Introducing Transformers
Boston biotech company Insilico Medicine appears to be the first to apply a new generation of AI, known as the Transformer model, to the drug discovery business. Back in 2019, researchers were wondering if they could be used to invent new drugs from biological and chemical data. Their first target was idiopathic pulmonary fibrosis, a lung disease.
They first trained the AI on a dataset related to this condition and discovered promising target proteins. A second AI then proposed a molecule that would attach to the protein and change its behavior, but would be less toxic or unstable. Human chemists then took over and created and tested the shortlisted molecules. They named their result lentsertib and recently completed a successful mid-stage clinical trial. The company said it took 18 months to arrive at a development candidate, compared to the four-and-a-half years it normally takes.
Insilico currently has a pipeline of more than 40 AI-developed drugs in development for diseases such as cancer, bowel disease, and kidney disease. And that approach is expanding. According to some projections, annual investment in this sector will increase from $3.8 billion to $15.2 billion between 2025 and 2030.
Partnerships between pharmaceutical companies and AI companies are also becoming more common. According to health intelligence company IQVIA, 12 deals were announced in 2024 with a total value of $10 billion. And last October, another pharmaceutical giant, Eli Lilly, announced a partnership with Nvidia, whose chips host the generative neural networks that transformer models rely on, to build the industry's most powerful supercomputers and speed drug discovery and development.
Given the strange economics of the pharmaceutical industry (the estimated failure rate of drugs entering clinical trials is 90%, and the cost of developing successful drugs is a whopping $2.8 billion), even small improvements in efficiency would yield significant benefits. Reports from across the industry suggest that AI is beginning to deliver these. The preclinical stage (the stage before human trials begin) has been shortened from three to five years to 12 to 18 months. And the accuracy has also improved. A study published in 2024 on the performance of molecules discovered by AI in early clinical trials found success rates of 80-90%. This compares to historical averages of 40-65%.
The design of new drugs typically begins by screening small organic molecules for promising biological activity and selecting those with the highest potential. AI can sift through libraries of tens of billions of molecules and use software emulations of these molecules to test properties such as potency, solubility, and toxicity. There is no need to place real molecules near the test tube. Jim Weatherall, one of the leaders of the effort at AstraZeneca, another major pharmaceutical company, said it has now culled the wheat from the chaff twice as fast as before, and more than 90% of the company's small molecule discovery pipeline is now AI-backed.
There are no errors when I try
AI can also help improve clinical trial design. One approach uses AI “agents” that behave as if they can think and reason. Back at GSK, head of AI Kim Branson demonstrated to our correspondents an agent-based system called Cogito Forge. When presented with a biological question, Cogito Forge can write its own code to help answer that question, collect the appropriate datasets, stitch them together, and create a presentation with graphs showing the conclusions drawn.
From there, one can generate a hypothesis about the disease that contains testable predictions and attempt to verify or disprove this through a literature search. Three agents are used for this search. One is to look for reasons why the hypothesis is appropriate, and the other is to look for reasons why the hypothesis is appropriate. Then look for reasons why it isn't. and the third decides which of the other two is correct.
Another area where AI holds promise is the selection of patients for clinical trials. Candidates' health records, biopsies, and body scans can be analyzed to identify those most likely to benefit from a new drug. Better participant selection means smaller trials, making them faster and cheaper.
But the most interesting use of AI to improve clinical trials is to create artificial patients (sometimes called digital twins) that serve as matched controls for real participants. To do this, the AI examines data from past trials and learns how to predict what would happen to a participant if he or she were to take its natural course without receiving treatment. Then, once volunteers are enrolled in the trial and given the drug, the AI creates “patients” with the same characteristics, such as age, weight, current condition, and stage of disease. The effectiveness of the drug in real patients is then measured compared to the progress of this hypothetical alternative drug.
If adopted, the use of synthetic patients could reduce the size of control groups in clinical trials and, in some cases, eliminate them altogether. It may also appeal to participants because it increases their chances of receiving treatment during the trial rather than being placed in the control group.
Modeling published in 2025 by San Francisco digital twin company Unlearn.AI suggests that this approach could have reduced control group size by 38% in an early Parkinson's disease trial and by 23% in another study on Alzheimer's disease. Additionally, early-stage trials typically lack control groups, and implementing these digitally could increase confidence in early signs of efficacy and improve how subsequent trials are designed.
Many proteins, molecules increasingly used as drugs but much larger than traditional drug molecules, have a tendency to wobble. This makes it difficult to determine the exact shape. The RNA molecules that form the basis of new types of vaccines are similarly tricky. And even more so with the complex membrane-based structures found inside cells. However, this is an area where understanding is rapidly advancing. AI is now being trained to model interactions between proteins and other molecules, predict RNA folding, and even simulate virtual cells.
Salt Lake City company Recursion has built an AI “factory” that photographs millions of human cells undergoing various chemical and genetic changes. This allows the AI to learn the patterns that connect genes and molecular pathways. And Owkin, a New York AI biotech company, trains its models on vast amounts of high-resolution molecular data from hospitalized patients. Orkin's boss, Tom Clausel, argues that by making discoveries that humans cannot, the research is moving toward true artificial general intelligence in biology.
These deviations from the usual tools of drug discovery raise the question of whether traditional pharmaceutical companies are at risk of disruption. In particular, OpenAI has made clear its expectations for models to reach high levels of competency in biology, training systems that can reason and discover in the life sciences. Currently, pharmaceutical companies have the advantage of a wealth of biological data and the context to understand and use it. Collaboration is the norm at the moment. For example, OpenAI is collaborating with RNA vaccine pioneer Moderna to accelerate the development of personalized cancer vaccines. But the balance of that advantage may change.
However, no matter who has the upper hand, if AI can elicit similar efficiencies in clinical trials, the odds of a molecule successfully making it through to clinical trials could rise from 5-10% to 9-18%. This may not sound like a big deal, but it means significant risk reduction for your business, and with it, drug development costs. In the medium term, this could lead to increased investment and an increase in the number of medicines brought to market. In the long term, if AI can solve biology, the technological possibilities for improving human health could be nearly limitless.
Curious about the world? For mind-enriching science coverage, sign up for Simply Science, our subscriber-only weekly newsletter.
