Q&A: Project Stargate and the viability of AI cancer vaccines

Vik Bajaj, managing director at venture capitalist firm Foresite Capital and CEO of Foresite Labs, which focuses on transforming healthcare through data-driven solutions, sat down with MobiHealthNews to discuss President Trump’s recently announced $500 billion Project Stargate and AI’s role in early cancer detection and vaccine development.

MobiHealthNews: You worked in the sector of early cancer detection at biotech company GRAIL. President Trump recently announced Project Stargate, which will use AI to detect and develop cancer vaccines. What are your thoughts on that announcement? Do you think the claims are feasible?

Vik Bajaj: The idea of detecting cancer early, which is something that we’ve worked on at GRAIL, is certainly doable, and you can look at GRAIL’s both analytical results, things that are mentioned and measured in the laboratory, as well as GRAIL’s clinical trial and – all of the work that the company has done. 

It leads to a conclusion that detecting cancer early is possible – and also that it is highly likely, if not assured that when you detect those cancers early, you will be able to do something about them – that really affects a dramatic change in patients’ lives, as compared to a scenario where those cancers were detected late, where they are disseminated and spread irreversibly to distant parts of the body.

So, that paradigm of early intervention, early detection, I think that’s established, not just by GRAIL, but by a host of other kinds of screening technologies, and now there’s an industry of players in the space that are working to extend that.

MHN: How would AI help scientists detect and develop a vaccine against cancer?

Bajaj: So, one question is detecting the cancer early, and I should mention even in the work that we did at GRAIL, AI is used to develop what’s called the classifier – the thing that takes in all of the input from these sequencing experiments that we do, the measurements of parts of the tumor that are shed into circulation, and associates that from all of these signals. Is this a cancer? And then, most importantly, where in the body is the cancer likely to be? Because otherwise you subject the patient to a real diagnostic odyssey. 

So, those conclusions, those diagnoses, depend on AI as it is, and then the question is, what do you do about it? Well, there’s a host of medical interventions. There are a number of companies that are developing cancer vaccines, and the two largest of those efforts, which have been tested in separate clinical studies in the U.K. recently, seek to develop what are called personalized cancer vaccines.

From a knowledge of the person’s tumor, you develop a vaccine which is likely to elicit the immune response of the body against that person’s tumor. So, how do you do that? Well, you have to determine what content to put in the vaccine, and that requires an understanding of what mutations are prevalent in the tumor, because you want to make sure that you’re targeting mutations that are prevalent, and ideally that they are in many, many or all of the cells in the cancer, rather than a few. And then also, of those prevalent mutations that are in many, many parts of the tumor, if not all, which ones are most likely to elicit an immune response.

And those are all problems where machine learning models of one sort or another have been used, as well as non-machine learning based approaches to tell you what’s more common, how has the tumor evolved over time, and then what of these targets that go into the personalized vaccine are likely to be more immunogenic or elicit the immune response that you want.

MHN: Is this something that could have happened without AI or is it happening faster because of AI?

Bajaj: There are elements of this that predate the modern version of AI that most people today think about when we just use the word “AI” in an unqualified way. So, in that sense, yes. But if you ask the broader question, could it have happened without collecting, for decades, huge amounts of data and then having AI approaches – maybe not even as modern as the ones that come today – interpret those data, I think the answer is no. It would be very unlikely. You could do bits and pieces of it.

We know, for example, that some tumors have mutations with very high prevalence. So, if you wanted to look for that one signal that you knew already was associated with a particular tumor, you could do that without AI. But to solve these big questions, how can you detect most cancers or all cancers in their earliest stages and know where it is in the body? How can you do that reliably with low false-positive rates? Those are things that depend on versions of AI. Absolutely. 

And as we get more and more data in the future, then our product development in the life sciences overall will be dependent on more and more sophisticated AI approaches, those that are fueled by larger and larger datasets, which only recently are coming into being.

MHN: So, because there’s so many different types of cancers, you would need different datasets for the different cancers. So, vaccines might work for one type of cancer but might not work for another.

Bajaj: It’s not just each type of cancer. Each type of cancer, different kinds of cancer have a lot in common. But another perspective is that each one is a very different disease, because its molecular drivers might be different. You know, the truth is somewhere in between. 

As I said, they have a lot in common, but they are also very different. They’re treated differently. Treated with different drugs that, in many cases these days, are targeted to the genetic rearrangements that are causing the cancer to emerge and grow in the first place.

And so, similarly, vaccines would be targeted and would be likely to work better in some cancers than others. And by the way, we don’t know that vaccines work yet. These are objects of promise, but need a lot of study to prove that they work – and those are very difficult and long experiments to do rigorously, also for the reason that you mentioned, that you would have to do that in a basket of related cancers, or cancer by cancer. So it’ll take a long time for us to say that that works.

MHN: When you heard about Project Stargate, what excited you, and what made you nervous?

Bajaj: The thing that excites me about it, just being very broad, in the last weeks we have had quite a bit of introspection in the U.S. about DeepSeek and U.S. competitiveness relative to China. 

I don’t think that it should be a surprise; I think what should surprise people is the amount of innovation happening elsewhere. When we take for granted that we are leaders in a space that maybe we even have a decades-long head start, that results in some complacency.

The reality is, it is not a decades-long head start. Progress in this field is so high that if you measure something today, and you measure it six months from now, it will seem like a discontinuous amount of progress. But actually, progress is just high. So, I would not be as alarmed about this as everyone is, but the lesson from the Stargate Project is that massive investments are needed to make resources available to a broader set of players to innovate. So, I think it’s good that they’re putting in those resources. 

And what worries me is that those resources are still concentrated in the hands of a few companies, academics, smaller companies and other groups that should be real contributors to the innovation sector in this field and have an activation energy or barrier. Not clear how to solve that, but it is clear that investments of that scale will be required to solve it and maintain our competitiveness.

MHN: Do you think this nervousness over DeepSeek is valid?

Bajaj: I think there’s many reasons to be nervous. It is not because this is a completely unexpected quantum jump in capabilities, or that it represents something that is years ahead, or that it represents a fundamental flaw in NVIDIA’s business model. I do not feel that any of those things are true.

I think, in the background, the reason to be worried is that there are competitors on a global scale, and this is a tight race between companies, between countries, between approaches that will be measured on a year-by-year basis.

MHN: Do you think it will also be measured on validation?

Bajaj: Depends on what you mean by validation. You know, scientists or in the healthcare space, we talk about validation, and it means something very rigorous. Every question, we put to the largest possible study to be sure that we are answering the question that we want, and proving something with as close to scientific certainty – at least about the accuracy – as possible.

So, validation in this space, it is not clear what it even means. There’s validation in the sense of model performance. In some segments of the AI universe, there are really well-structured tests that can be used to compare models. In other segments, there aren’t. So it’s hard to talk about validation in a general sense.

By the way, we don’t seek that level of validation and important experiences in everyday life, right? We don’t validate drivers in the way that driverless cars are validated. We do not validate doctors in the way that AI expert systems that are reasoning about medicine are validated. We just have to have that perspective as well.