Artificial intelligence (AI) is rapidly becoming a central tool in cancer research and care. From the 2024 Nobel Prize for AI-based protein design to the rollout of a new AI tool at the FDA to accelerate reviews and other agency processes, the technology is reshaping drug discovery and diagnostics. In context of the recent oncology-focused research meetings — the 2025 Annual Meeting of the American Association for Cancer Research in April and the American Society of Clinical Oncology 2025 in June — DDN asked leading cancer researchers how AI is changing drug development, treatment, and diagnosis.
How will AI affect cancer treatment and diagnosis?
Credit: ThirdLaw Molecular
Christian Schafmeister, Founder and President of ThirdLaw Molecular
AI is becoming a critical tool for accelerating discovery, but at ThirdLaw Molecular, we see the greatest near-term impact in smarter molecular design and biomarker discovery. AI helps us prioritize targets, predict binding interactions, and simulate molecular behavior — dramatically reducing trial and error in early discovery. For treatment, AI-driven companion diagnostics will be key for matching therapies to patient subgroups more precisely and improving clinical trial success rates. However, the value lies not just in prediction, but in validation, and that’s why we combine AI insights with rigorous chemistry and biological testing to ensure clinical relevance.
Credit: Laura Shachmut (Shachmut Photography)
Catherine Sabatos-Peyton, Chief Executive Officer at Larkspur Biosciences
The potential for AI in cancer drug discovery is huge, but I think we have a ways to go from where we are now to how that potential gets borne out. AI has already been employed in the drug discovery chemistry side of things for a long time and can only continue to build from there, which I think is amazing and has a lot of potential. We tend to think about these questions: How can we use it to better translate medicine into the right precision patient populations for cancer? How can we really understand which patients are most likely to respond and get medicines to those patients most rapidly?
Then, there's also the biology: How can we use AI to inform target discovery? Does it help us understand new targets or, perhaps, even known targets that might be important? So much of that depends upon the data that's available and asking the right questions of the tools, which I think, today, still requires a human brain. I think the potential of AI is huge, but I think there remains a significant gap between what we're able to do right now and what ultimately will be possible. It's a super exciting one to bridge, and I'm excited to see the work that people are doing.
Credit: City of Hope
David Craig, bioinformatician and translational scientist at City of Hope
One of the biggest challenges in clinical trials is sifting through patient electronic medical records for clinical trial recruitment. How do we identify patients, and how do we structure them? One of the really amazing things that a lot of us are doing is using these large language models (LLMs) to structure our data. Previously, we would have people with some domain knowledge who understand the databases to go in and read the doctor's notes and put them in context. In many ways, now we can automate that, and we're at the early stages where we're starting to train the LLMs and make sure they do everything right. For example, does the LLM get the stage right? Do I have patients with a KRAS-G12C stage three or stage four cancer?
Another part of AI, which I think is really interesting and is spurring a revolution in digital pathology, is how the models are built. The first layer is trained off of everyday images — pictures of dogs, pictures of cats, pictures of animals. They're being categorized, and that’s the foundational model. Then, we can take that same model and train it with histopathology images, and we see that it does a better job. If you think about it, it's crazy. How can training on picking out Chihuahuas from Great Danes do that? But if you think about an image, there's points where the human eye will see a transition from colors or edges. It comes down to frequencies and light, and that becomes a starting point. That allows us to not to have to start from scratch. The past paradigm would be to just use histopathology from the get-go, but then we would worry about how many millions and billions of images we would need to train the model. Now, the first step is to take a pre-trained foundational model, which was trained on a lot of different things, and adapt it because it's already learned certain principles.
If we bring it back to drug discovery, I think digital pathology is going to see a lot of interesting advances. For example, for tumor-infiltrating lymphocytes, can we quantify those? Can we annotate those? Those are things that a lot of folks are working on. There may be other types of complex structures within a histopathology image that could be informative.
Credit: Signios Bio
Marco Corbo, Senior Bioinformatics Scientist and Director of Bioinformatics at Signios Biosciences
Machine learning approaches have existed for the past 20 years. The difference nowadays is that with the number of data types that we have, we can train those models to give us even better insights. The bottleneck of all these models is that they are only as good as their underlying training model. That means that if we want to find a genetic variant with the most relevance in a specific phenotype, we need to have a trained model with millions of variants from people from all over the world to understand which variant might be the most affected. Now, as researchers collect all of these data in databases, with the increasing computational power of different cloud-based platforms, and through integration with multiomics tools, we can train these models even better. Now, these AI and machine learning tools can predict specific phenotypes based on very few data points.
In terms of drug discovery, AI tools can also save us time. Before, we would create hundreds of compounds based on one that we already had and start adding modifications to those compounds. Then we would have to test those compounds. But nowadays, with AI and machine learning tools, we don't need to screen hundreds of markers or potential putative drugs. We can just narrow it down to a few. These drug candidates and markers can also be completely de novo. With these models, we can design a drug to target a specific receptor on the surface of a T or B cell without having any previous information on any other drug that was targeting that receptor. That's the power. We can really reduce the time for the first steps of drug design. That will, of course, help get a new drug into clinical trials much sooner.
Responses have been edited for length and clarity.
