Artificial intelligence is changing how researchers search for medicines. It can help identify promising targets, design candidate molecules and find existing drugs that may suit particular patients. But these systems cannot establish on their own that a treatment works in people.
Testing more options for individual patients
One example comes from a trial at the Medical University of Vienna, where doctors were treating Paul, an 82-year-old with an aggressive blood cancer. Six courses of chemotherapy had not eliminated it, and the usual cancer drugs had not worked.
Researchers tested a tissue sample containing both cancerous and normal cells. They divided it into more than a hundred pieces, exposed those pieces to different drug combinations, and used robotic automation and computer vision to track the cells’ responses. This let the team assess many treatments at once instead of waiting through months-long treatment courses.
The leading candidate was unsuitable because Paul was too frail to take it. He received the runner-up, a cancer drug marketed by Johnson & Johnson that his doctors had not tried because earlier trials suggested it would not help his type of cancer. Two years later, he was in complete remission.
The case illustrates how testing patient tissue could help clinicians compare options when standard treatments have failed. It does not mean every patient can be matched to a successful drug this way: the result depended on testing the sample, and the top-ranked option was not appropriate for Paul.
Where machine learning fits in drug development
Drug development still follows a familiar sequence: researchers choose a target in the body, design a molecule to affect it, make and test that molecule in a lab, and then study it in people for safety and effectiveness. Traditional screening involves repeated rounds of testing and adjusting candidate molecules. Many that appear promising in lab conditions fail when tested in humans.
Machine-learning models can use large collections of chemical, biological and scientific data to estimate how candidate drugs might behave. Researchers can use those predictions to screen out some weak options before investing in lab work. That can make early exploration broader and faster, while directing experiments toward candidates with better prospects.
Some teams use natural-language processing to search scientific literature for possible drug targets. Information from papers and gene sequences can be organized into knowledge graphs, which represent connections such as one factor causing another. Models can then suggest targets worth investigating, including connections that researchers may not have found by reading the literature manually.
Other companies focus on the molecules themselves. Generate Biomedicines is using generative AI to propose protein structures with chosen properties. Absci uses machine learning to redesign parts of existing antibodies, then synthesizes and tests the strongest designs. These approaches aim to explore candidate structures that would be difficult to examine one by one.
A larger search, with limits
The potential scale of the search is one reason the technology has drawn investment. The source describes a vast gap between the molecules already considered by drug companies and the much larger range of structures that could be possible under the rules of organic chemistry. Computational tools may help researchers explore more of that space and select candidates for further testing.
AI-assisted drug discovery is also moving into clinical trials. The source reports that two drugs developed or co-developed by Exscientia had started the clinical-trial process since 2021, with two more on the way to submission. Around two dozen drugs developed with AI assistance were in or entering trials at the time described.
Those milestones show that AI-designed treatments are being evaluated, not that they have been proven successful. Predictions about how a molecule behaves in a body can be wrong. Lab experiments and human testing remain necessary, and the slow, expensive parts of drug development cannot simply be removed.
What changes for patients and researchers
For patients, better matching could reduce exposure to treatments unlikely to help, and a wider range of candidate medicines could eventually offer options for diseases that remain difficult to treat. The Vienna trial points to one way of testing a patient’s cells against multiple drugs, while AI drug design aims to create new candidates in the first place.
For researchers, these tools shift some effort toward data, prediction and prioritization. They can help decide which targets and molecules deserve experiments, but people still need to judge whether a predicted connection makes biological sense and confirm results in the lab.
The near-term promise is a more efficient search, not instant certainty. AI can help researchers move through more possibilities and focus their resources. Whether any particular candidate becomes a safe, effective medicine still depends on evidence from experiments and clinical trials.