Patrick Schwab is not your typical pharmaceutical researcher and his workplace is not your typical pharmaceutical laboratory. There are no benches and no bubbling liquids. White lab coats are also absent. Instead, Dr. Schwab is dressed all in black. But it is appropriate attire for a man whose workplace is in King’s Cross, an area that was once railway yards and industrial buildings but is now, after a makeover, one of London’s most fashionable districts.
Dr. Schwab works for GSK pharmaceutical company. Their work aims to reimagine the future of drug manufacturing using the same modern branch of computer science, artificial intelligence (AI). He is applying this to transfer as much of the load as possible from glassware to computers: in silico drug design rather than in vitro.
For this, he is developing a software tool called Phenformer, which he is training to read the genome. By combining genomic information with phenotypes – the biological term for the physical and behavioral outcomes of particular genetic combinations – Phenformer learns how genes drive disease. This allows it to generate novel hypotheses about diseases and their underlying mechanisms.
meet the transformer
Insilico Medicine, a Boston-based biotech firm, appears to be the first company to apply a new generation of AI, known as the Transformer model, to the business of finding drugs. In 2019 its researchers wondered whether they could use biological and chemical data to invent new drugs. His first target was idiopathic pulmonary fibrosis, a lung disease.
They began training the AI on datasets related to this condition and found a promising target protein. Then a second AI suggested molecules that would capture that protein and change its behavior, but would not be too toxic or unstable. Human chemists then took over the task of creating and testing the shortlisted molecules. They called the result rentosertib, and it recently completed successful mid-stage clinical trials. The firm says it took 18 months to reach a development candidate – compared to the typical timeline of four and a half years.
Insilico now has a pipeline of over 40 AI-developed medicines it is developing for conditions such as cancer, gut and kidney disease. And its approach is spreading. One estimate suggests that annual investment in the sector will increase from $3.8 billion to $15.2 billion between 2025 and 2030.
Collaborations between pharma companies and AI companies are also becoming common. A dozen deals were announced in 2024, with a total value of $10bn, according to health-intelligence company IQVIA. And last October Eli Lilly, another pharma giant, announced a collaboration with Nvidia, the firm whose chips host the generic neural networks on which Transformer models rely, to build the industry’s most powerful supercomputer, and thus speed up drug discovery and development.
Looking at the pharmaceutical industry weird economics-The failure rate of drugs entering clinical trials is 90%, pushing the cost of developing a successful drug to $2.8 billion – even small improvements in efficiency would provide big benefits. Reports from across the industry show that AI has begun to deliver these. This has reduced the preclinical phase (before human trials begin) from three to five years to 12–18 months. And this has improved the hit rate. A study published in 2024 on the performance of AI-discovered molecules in early clinical trials found an 80-90% success rate. This compares to a historical average of 40–65%.
The design of a new drug typically begins by screening small organic molecules for promising biological activity and selecting the most probable. AI can sift through libraries of billions of molecules, testing properties like potency, solubility, and toxicity using software simulations of those molecules, eliminating the need to get anywhere near real molecules in a test tube. Jim Wetherall, who is in charge of this activity at AstraZeneca, another big pharmaceutical company, says it separates the wheat from the chaff twice as fast as before, and more than 90% of the firm’s small-molecule discovery pipeline is now AI-assisted.
test and no error
AI is also helping to improve test design. One approach involves AI “agents” that behave as if they can think and reason. Back at GSK, head of AI Kim Branson gave your correspondent a demonstration of an agent-based system called Cogito Forge. When asked a question about biology, Cogito Forge is able to write its own code to help answer that question, gather appropriate datasets, stick them together, and then create a presentation – complete with charts showing the conclusions it has drawn.
From there it can generate a hypothesis about a disease, including testable predictions, and attempt to verify or falsify it with a literature search. That search employs three agents: one to look for reasons why the hypothesis is good; A second to look for reasons why this is not the case; And a third will decide which of the other two is right.
Another area where AI is promising is in selecting patients for trials. It can analyze candidates’ health records, biopsies and body scans to identify who might benefit most from a new drug. Better choice of participants means smaller—and thus faster and cheaper—trials.
However, the most interesting use of AI to improve trials is the creation of synthetic patients (sometimes called digital twins) to act as matched controls for real participants. To do this the AI goes through data from previous trials and learns to predict what might happen to a participant if they follow the natural course of their condition instead of seeking treatment. Then, when a volunteer is enrolled in a trial and given a drug the AI creates a “patient” with the same characteristics, such as age, weight, existing conditions and disease stage. The efficacy of the drug in the real patient is measured against the progress of this virtual option.
If adopted, the use of synthetic patients would reduce the size of control arms of trials and could potentially, in some cases, eliminate them altogether. They may also attract participants, as they will be more likely to receive the treatment under trial rather than being placed in a control group.
Modeling published in 2025 by Unlearn.AI, a digital-twins firm in San Francisco, suggested this approach could reduce the size of the control arm by 38% in an early trial of Parkinson’s disease and by 23% in a separate study on Alzheimer’s disease. Additionally, early-stage trials in general, which sometimes lack a control arm, can now be presented digitally to increase confidence in early signals of efficacy, and improve how subsequent trials are designed.
Many proteins – molecules rapidly deployed as drugs but which are much larger than traditional drug molecules – have a tendency to move around. This makes it difficult to determine their exact shapes. RNA molecules, the basis of a new class of vaccines, are equally intriguing. and the complex membrane-based structures found in the interior of cells, and much more. But this is an area where understanding is advancing rapidly. AIS now being trained Modeling interactions between proteins and other molecules, predicting RNA folding, and even simulating virtual cells.
Recurson, a Salt Lake City firm, has created an AI “factory” that features millions of human cells undergoing various chemical and genetic changes. This allows AI to learn patterns linking genes and molecular pathways. And Ovkin, an AI biotech in New York, is training its models on huge sets of high-resolution molecular data from hospital patients. Ovkin’s boss Tom Clozel argues that by making discoveries that humans cannot make, the work is moving toward true artificial general intelligence in biology.
These deviations from the usual tools of pharmaceutical discovery raise the question whether traditional pharma companies are at risk of disruption. OpenAI, in particular, has been clear about its expectation that models will reach high levels of capability in biology, and that it is training systems that can reason and make discoveries in the life sciences. Currently, pharmaceutical companies have the advantage of a wealth of biological data and the context to understand and use it. For now, collaboration is the order of the day. For example, OpenAI is working with RNA vaccines pioneer Moderna to accelerate the development of personalized cancer vaccines. But that balance of benefits could change.
However, whoever gains the upper hand, if AI can achieve similar efficiency gains from clinical trials then the chances of a molecule successfully navigating the clinical-trial journey could increase from 5-10% to 9-18%. This may not seem large, but it represents a massive reduction in risk to the business, with a concomitant reduction in drug development costs. In the medium term, this could lead to an increase in investment and the number of medicines coming to market. In the long term – if AI can solve biology – the technological possibilities for improving human health could be almost limitless.
Curious about the world? To enjoy our mind-expanding science coverage, sign up just scienceOur weekly subscriber-only newsletter.
