The Stanford professor talks about the extraordinary potential of AI and why the human-machine partnership is where the real power lies.
Garry Nolan, the Rachford and Carlota A. Harris professor in the department of pathology at Stanford University School of Medicine, has published more than 350 research articles, is the holder of 50 US patents, and has been honoured as one of the top 25 inventors at Stanford University. In short, there are few people who understand cancer research as well as he does. Here, he talks to Healthcare Today about where artificial intelligence (AI) can help with developing treatments for cancers, how it works and what its limits might be.
You’ve said there’s no golden bullet, that each cancer needs its own treatment strategy. How does AI help us map that complexity and identify the right treatment pathways?
Each cancer originates from a different organ, and every organ is its own ecosystem. That’s why, although there are some common features across cancers, each type often has its own distinct origin and trajectory.
Because of this complexity, the bluntest and most widely used tools we have are still chemotherapy and radiation, both designed to stop cell division entirely. But that approach also harms healthy cells that still need to divide, which is why patients experience side effects like hair loss.
This is where AI has extraordinary potential. No single human can hold in their head the millions of studies, data points and pathways involved in all the different cancers. Large language models, however, can scan vast literatures and connect concepts that might otherwise remain buried in a footnote. A new layer of agentic AI makes this even more powerful: like a supervisor directing a dozen interns, it can send multiple processes into the literature, retrieve and organise the most relevant insights, and present them in direct response to a clinician’s question.
“AI will help us solve cancer, accelerate fusion energy and do many other great things.”
There’s a lot of hype around AI “curing cancer”. How much of this is marketing spin, and how much is real scientific progress?
What AI really changes is speed – and the ability to marshal lots of small, focused searches at once. I think we’ll be having a very different conversation in five or ten years as agentic systems become far more sophisticated.
In my lab, we’ve already built an agentic-AI pipeline to tackle the tumour-immune interface – that bewilderingly complex dance between what the immune system is trying to do and what the cancer is trying to prevent. We can now take spatial profiling and other raw data, throw it into the system, and get out structured models of what the tissue ecosystem might look like. The AI organised thousands of questions, hypotheses and possible laboratory tests and, when fed real biopsy data, was able to spot patterns that would have been impossible to find by hand.
Take tertiary lymphoid structures, for example. These are lymph-node-like formations that sometimes appear inside tumours – forward camps the immune system sets up to fight the cancer. The more of these structures a tumour contains, the better the outcome after therapy. Our agentic system was able to trace back, in hundreds of biopsies, to the originating cells that seed those structures. That is a basic-science discovery of huge potential: it suggests we can look in other cancers for the early, aborted versions of these structures, ask why they were halted, and then either prevent the block or encourage more seeding. Those are precisely the kinds of therapeutic footholds you wouldn’t find without computers trawling millions of correlated data points.
The difference between last year and this year in what the AI can do is enormous. Yes, the headlines can be clickbait, but my view remains optimistic. AI will help us solve cancer, accelerate fusion energy and do many other great things. Crucially, though, humans must stay in the loop to guide the questions and interpret the answers. That human-machine partnership is where the real power lies.
Will AI get us to develop cures for most cancers in a piecemeal fashion – cancer by cancer – or are we heading towards a tipping point where progress becomes exponential?
I don’t think there will ever be a single drug that cures all cancers. Each cancer is too different. But I do believe there will be a small set of immunotherapies that can apply across many cancers.
Of course, therapies must also be tailored. Solid tumours are very different from leukaemias, which flood the blood system, or from cancers like ovarian, which can exist in both solid and free-floating phases. A therapy that works by creating lymphoid structures in a solid tumour won’t help much against acute myelogenous leukaemia, for example, which needs another approach.
What excites me about AI is that it opens entirely new doors to pursue these possibilities. I can honestly say I’ve never been more excited about what is possible.
For instance, I think we need new kinds of atomic imaging to understand molecular structures. AI is helping me design such an instrument with a company I’ve started in San Jose. The point is that the real code of life isn’t only in the Watson-Crick letters of DNA. It’s also in the 3D shape of chromosomes. Each cell has a different DNA architecture, and it’s the cell’s ability to read that shape – not just the sequence – that explains why cells behave differently. That’s where I think AI can help unlock the next level of understanding.
“Large language models hallucinate, but that’s not always a flaw – it’s a feature.”
Is it fair to say the AI is being used, not just as an accelerant for existing approaches, but it’s opening up entirely new therapeutic frontiers?
Cancer research has already taken most of the low-hanging fruit. The next step is to understand not just how a single cell makes decisions, but how cells interact with each other in complex systems – the immune system or the tumour microenvironment – which may look chaotic, but are actually highly organised.
AI can help us identify the critical points in these systems. Think of it like finding the maxima or minima of an equation: if you can pinpoint the one or two pressure points, targeting those could make the entire system collapse. No individual scientist knows every cell type or every pathway – the immune system alone looks like a map of the universe, with endless cell populations and interactions. Experts master single niches, but only AI can hold the totality of that knowledge at once and return answers when you ask the right questions.
What’s striking is that AI isn’t just a tireless assistant. It can also be creative. Large language models hallucinate, but that’s not always a flaw – it’s a feature. By making slightly noisy, unexpected connections, they can surface ideas we wouldn’t otherwise see. Creativity itself often sits at the edge of instability: some of history’s most original minds had tendencies toward schizophrenia, while autistic individuals often show extraordinary focus. In a sense, AI combines both traits – the ability to hallucinate just enough to generate new ideas, while also focusing relentlessly on detail.
That’s how discoveries happen: by making unexpected correlations, opening doors no one realised were there.
We’ve talked about the advantages of AI and what it can do, but what can’t AI do yet in cancer research?
The key with AI is that it doesn’t do anything until you start it. You have to ask the question that triggers the outcome. That’s why I don’t see it as some all-knowing genius about to surpass humanity in the next couple of years. Instead, I think of it as multiple specialised agents – each expert in a narrow area – with a supervisory layer on top that knows which one to call on.
For example, I was looking this morning at a volume of research papers from 60 years ago – more than a dozen articles full of interesting data. I’ll use AI not just to consolidate them, but to structure the question so it finds what’s unique in each paper, not only the commonalities. Those one-off insights, the tidbits buried in just a single study, can sometimes be the most valuable.
I see AI as an organiser and accelerator – a way to uncover relationships that no one would have suspected, because it can move through data so quickly.
How should we balance the excitement of AI-driven breakthroughs with the realities of clinical trials, regulation and patient safety?
At the end of the day, you still need to test in humans. I’m genuinely conflicted about the use of animals – I hate it, but I also understand its value. Where AI is really helping is in designing the clinical trial itself. That’s enormously complex: you have to navigate local, national, or even international regulations, secure Institutional Review Board approval, and meet requirements of which you might not even be aware.
AI can act almost like a compliance co-pilot. You can say, “Here’s what I’m trying to do – am I in line with the rules?” and get an answer in minutes. Something that might once have taken six months can now be done in 15 minutes. It can also check whether a trial is statistically powered correctly, and even support adaptive designs, such as Bayesian trials that evolve as results come in.
In the past, I would have had to deploy a team of statisticians to make sure we weren’t overfitting the data or drawing shaky conclusions. Now AI can handle much of that heavy lifting. I still take everything back to a statistician for review and sign-off, but at least I’m not asking them to build it all from scratch.
