AI in the lab: from tweezers to spoons

In this episode, we talk with Dr. Obi Griffith from 
Washington University about how AI is currently impacting cancer research in the lab and how it may influence the future of cancer research. 

Transcript

Copy

 CARRIE: Welcome back to Cracking the Cancer Code, a podcast series about the researchers who use data to fight cancer. I'm Dr. Carrie Wright, a senior staff scientist at the Fred Hutchinson Cancer Center. I'm the head of content development for the ITCR Training Network, a collaborative effort funded by the National Cancer Institute of researchers around the United States aimed at supporting cancer informatics and data science training. 

CARRIE: Since our last episode, the cancer research landscape has changed dramatically, and a big reason for that is AI. AI and large language models are everywhere and impossible to escape, and this includes the research lab. We wanted to bring you an episode that focuses on how widespread adoption has changed what life is like in the cancer informatics research lab, as well as what AI can and cannot do in the fight against cancer. Today, we're joined for this discussion by Dr. Obi Griffith, a researcher at Washington University in St. Louis and one of the co-PIs for the ITCR Training Network 

OBI: Hi guys, I'm Obi Griffith. I work at Washington University. I lead a cancer genomics and cancer informatics lab. Uh, we're really focused on kind of interface between genomics and precision oncology. Uh, we develop open source tools to try to facilitate the application of genomic methods to, um, improving patient outcomes for cancer patients. And that is focused around several large efforts, um, including things like clinical interpretation of variants in cancer knowledge base, and the drug gene interaction database, and, um, other tools that are increasingly using AI to think about ways to parse knowledge from the literature and make clinical decision-making more efficient, let's say, in, in cancer research and cancer practice. 

CARRIE: So Obi and I have actually talked a lot about AI use in the lab, and I know, Obi, that you have some really interesting and unusual uses that I think our audience would be excited to hear about. So could you please tell us how your lab uses AI?  

OBI: Yeah. Um, I feel like this rollout of LLMs and agentic AI has changed, like, the daily life as a researcher much more than it's changed anything kind of fundamental in the research or certainly clinical space yet, but hopefully we'll get there. Um, but for sure, just personally, and many people in the lab are using it in so many different ways from morning till night, both personal and, and research applications. Um, so we have been playing a lot with, um, for example, Google's NotebookLM, and have been really appreciating how useful that is to kind of scope your typical AI chatbot experience to a specific problem. So you can upload, like, a set of related papers and get it to summarize, like, an area of the literature maybe that you're less familiar with. And we even did use it to kind of generate a podcast that gave a high-level summary of a new area. It was about, um- parathyroid adenomas and carcinomas, which is not something we'd previously researched. And to kind of identify some gaps in the field that, that might be worth pursuing and share that with our collaborators. And we heard from them that they really enjoyed like sitting around the, the speaker in the lab, so to speak, and listening to kind of summary of their research field Um, and where some gaps might be. And it, it-- I feel like it really did help us kind of crystallize on what the collaboration would focus on over the next, like, weeks and months. Uh, but we've also been using NotebookLM in a bunch of other ways. Like, I've been using it, um, in one of my daily tasks as an editor for the Civic Knowledge Base to fact-check submissions. So you can take, like, an evidence item that someone submitted, and you can upload that along with the, the source publication for that evidence and ask it to sort of systematically fact-check every fact and detail and sentence in that submission. And, um, it does a surprisingly good job. It doesn't always catch everything, but so far my experience is that what it catches is good. Like, it'll say, "Oh, this sentence is a little bit misleading," or, "It looks like someone mistyped and put the wrong confidence interval for this, um, odds ratio," or whatever. And it has, like, a pretty high success rate at identifying little, little problems like that, um, some of which the human eye will actually struggle to, to find, I think. As a PI, I use it in a lot of ways to try to tackle some of those, like, tedious, mundane problems, like filling out, you know, required forms for grant submissions and, and the like. I've also been so far unsuccessfully playing with, um, building my own agent to try to handle my calendar scheduling problems. Um, so those are like just some examples. But then we're also, as you, as you alluded to, really exploring the possibilities of integrating it into some of our tools and some of our research questions. 

CARRIE: So there's been some examples lately and stories in the news about people trying to use public AI tools to help treat their pets who have cancer. Uh, and there was a case of someone in, in New Zealand who used, I believe it was ChatGPT and AlphaFold to, uh, figure out a vaccine for his dog's cancer. So what are your thoughts on that? 

OBI: Yeah, I mean, this is very close to home for us because we, um... A major focus of the lab, which I didn't mention before, is actually designing cancer vaccines or neoantigen vaccines. So the idea here is that you sequence a cancer, sequence a tumor, and you identify mutations that are specific to the tumor, and you predict peptides that would result from those mutations, and you use a bunch of algorithms to predict which of those are likely to kind of stimulate the patient's immune system the most. And you kind of-- You design a vaccine, it could be mRNA or DNA or protein-based platforms. You put that in the patient, and it's supposed to kind of stimulate the patient's immune system and kind of remind them that this, this cancer, this tumor with these little pieces of protein are, are foreign. They're not supposed to be there, and that you, you should attack that and get rid of it from the body. And so we've supported, you know, at this point, probably more than a dozen trials in humans, um, for this approach and, and have actually also started a small trial and treated one dog with a cancer vaccine. So we were super interested to see that report of someone basically doing it themselves. And, you know, it's hard to know because it's reported through the media, but at, like, at a high level, it sounds like they're doing something reasonable. I think the devil is in the details. I would love to, like, dig into their exact design. Um, but someone shortly after that, inspired by that, released this thing called, I think they called it Cancer Studio, which is a really slick-looking app that they, I guess, like, vibe coded up, um, that kind of has a full pipeline, like upload your sequence data and get out at the other end a vaccine design that they kind of low-key encourage you to, like, send off to a manufacturer, a GMP or otherwise manufacturer, and you'll have a vaccine to give your dog or cat. Um, which is probably not recommended to just do that on your own.

CARRIE: When we are talking about pet sequencing, you know, we can use public tools without any problem because we're not really worried about the data privacy of our pets, but that's a challenge that we do have for humans. 

OBI: I, I think it's really illustrative of where we are with this, like, AI revolution, like both what's empowering about it, that you got-- you can use these agentic coding systems to whip up like a tool that's that impressive so quickly just based on the inspiration from also someone that's kind of doing something a little bit crazy, like garage band style vaccine designs. Um, but that's definitely the world we're living in right now, and I have very conflicted feelings about it. On the one hand, like I wish our pipeline and our, you know, we have graphical interfaces that help us do our vaccine designs. I wish it looked as slick as the one this guy vibe coded up over a couple weekends. So it's making me think maybe, maybe we should like up our game in terms of our tools and our design for that. But on the other hand, I know about all of the like complexities of the process, and I question whether like every one of those complexities is being accounted for in something that was put together sort of this quickly.  ​ Yeah. 

CARRIE: So I know that you and I also have some concerns for usage, particularly when it comes to trainees , about understanding the decisions that we make in our software development or our analyses. Can you talk about your thoughts on this? 

OBI: Yeah. Uh, my enthusiasm or, and, or, uh, cynicism or concern about AI, the pendulum swings between those ends so wildly from moment to moment, even during the day. Um, but I guess, like, the one observation that you hear a lot of people making in the software development field, which I think is where this is having the, the, the, the biggest impact soonest, is something along the lines of, um, you know, s- more senior developers who have really spent like years or decades kind of developing the experience and the critical thinking about code. It's like a real amplifier for them pretty quickly because they can kind of-- There's some maybe tasks that they haven't gotten to that they, they have a very solid understanding of what needs to be done, and there's just not enough hours in the day, so they get an agent working on it and, and they can vet what it does quite accurately and quickly. And so it's super empowering for them. But then at the other end of the spectrum, there's, like, a general sense of concern about the, maybe like the junior developer who's just learning. They haven't developed that intuition or experience, and they're learning a, basically a new way of coding. Um, and so like, yeah, we'll-- There's, there's constantly these predictions that like five years from now, we'll all be like standing amongst a pile of crumbling technical debt-laden tools that were vibe coded up by developers over the last several years, and now no one has any clue how they work. Um, but it's-- I don't know. It's so hard to know, like, is that a real concern or not? Is that really any different than the 43 levels of abstraction we're already standing on? Um, and do we discount the new ways of thinking that these junior developers working in a kind of cyborg mode with AI are gonna develop that, like, they will look back on us and these quaint attitudes, like, with great humor in 10 years? Like, a-any of these things are possible in my mind. Um, for this-- I was just gonna say, on the science side, I, I guess I'm optimistic that the phenotype of the scientist is generally super questioning, wanting to understand how things work. That's part of the point and why you become a scientist. And so that there is, like, a strong instinct to continue to, like, dive into things, take them apart, and understand how they work, and not just, like, vibe code up an experiment and assume that the results are correct. I guess that's my optimism for the science side. 

CARRIE: Yeah, it's really hard to know exactly how this is going to impact our work and our skills, especially for trainees trying to acquire new skills, and I definitely share your concerns. There are some studies actually that suggest the way trainees use AI to learn can really impact their uptake of skills, in particular for programming. So if they just ask AI to give them what they need for a task, they aren't going to learn very well. But if they give part of the code and they ask for specific improvements, or they try to have help from AI in terms of understanding the code, studies suggest that that might really assist them in learning with AI. But we really need longer term studies to understand how that impacts their upskilling. And then further, there's a lot to be learned in terms of de-skilling. So if someone, you know, maybe learned some programming but they're not doing it very often and they rely on AI a lot, are they going to lose those skills more rapidly? And there are some studies that suggest that is the case.

OBI: Yeah Yeah, I have those concerns, but I also, I guess I'm, I'm excited about the possibility of tackling some of the backlog of ideas or plans that in theory AI could, could empower and we're, we're trying to do some of that 

CARRIE: Yeah, there are also some really great positives right now in terms of AI helping reduce our burden with tedious tasks. It, science has a lot of tedious tasks. Can you discuss what you're most excited about in terms of this? 

OBI: I think, what... There's a long list of things I'm the most excited about, but, like, one obvious one for our lab is focused around this problem of variant interpretation. So we've been leading these efforts to build a public knowledge base of variant interpretation for many years now, and it is slow going. I mean, it's just, like, incredibly laborious manual effort that requires, like, PhD or MD, PhD level experts to stop the 100 things they're doing and dedicate a bunch of time to, like, doing a deep dive into the literature on a certain variant or topic and extract, painfully extracting and synthesizing the knowledge and submitting it to this knowledge base. And then other experts have to come and vet all of that to get to this hopefully high quality variant interpretation product. And, you know, since we started almost 20 years ago, there's always been this dream that somehow eventually we will have text mining or other NLP or AI approaches that will help scale this, like, truly unscalable problem, right? Where the literature every year gets, like, an order of magnitude bigger than, than the number of kind of quirky bio curators that exist that are willing to try to, to extract knowledge. So it's always been this dream, but every time we've tried it, like every few years, we would attempt to roll out some new text mining or NLP approach to scale some part of the problem and generally be dissatisfied. Like, on the research side, we'd be like, "Oh, that's kind of cool that it can identify genes from papers," but it would never really move the needle in terms of the actual bottlenecks of, of effort. Like, at the end of the day, you still needed to read the paper and extract the knowledge from it, and having a few little parts of that automated didn't really help that much. It sort of feels like now we're finally, we finally have the technology that might possibly work for the first time at this task. I mean, we, we haven't proven that it can, but more than ever it feels like the possibility actually exists that we could start to really scale this effort and, and automate major parts of it. 

CARRIE: Yeah, it's really wild how text mining has dramatically changed and that we're getting much closer to more accurately interpreting larger sets of text, like a full manuscript So there's also some really wild claims about how AI can basically replace us as researchers, and there's also a history of some major AI failures in the past. So what do you think about all of that? 

OBI: Yeah, like maybe, uh, I don't know if you remember IBM Watson, like the excitement around that, and I'm sure like IBM gained some useful insights which contributed to the develops- development of some, some novel methods and stuff. But it-- let's be honest, it really did not pan out the way it was pitched. And maybe something like that now could, although I feel like we are also in like a similar hype cycle where you have to treat like almost all the claims you hear with extreme credulity I mean, every time I hear, um, one of these tech bros say, "Good news, cancer will be cured in five years," it just makes my head explode. And it's sort of like everything you said up to this point I have to now completely discount. And so I don't know if any of your enthusiasm is warranted because I'm enough of an expert in this one area to feel like 99.9999% confident that you're wrong about that And then it, it makes me wonder, is there like-- is there an expert in all the other areas that they're talking about? Like, is there some lawyer somewhere that's like, "Yeah, no, what you said about like lawyers all going away, that's not true either." It's just that I don't have the expertise in those things to be able to evaluate how silly each of those statements are. Or, or is like cancer and data science like maybe a slightly more uniquely difficult problem to solve by AI? I suspect it's the former, though, that like all of the claims are overstated.

CARRIE: Yeah, I know. There are always these wild claims about the cure for cancer from someone who typically doesn't have a biomedical background.

OBI: Right. Right. 

CARRIE: Yeah, it's hard to say about other fields, but indeed for cancer, I agree that it's more likely going to be incremental steps to help with individual cancers instead of this one giant leap for all cancers, and it's all, it's all taken care of by AI. 

OBI: Mm-hmm. 

CARRIE: So if you had one overall message about AI and cancer research for those people who've heard these wild claims, what would it be? 

OBI: Um, I don't know. I guess I would say like the sort of overall message is that there's, you know, there's tremendous potential, I think, for cancer research to be advanced by AI, but at the same time we should moderate our expectations because the scale of the problem is such that... I've been trying to think of like actually asking AI to help me make a cartoon to depict this, and I'm imagining a scientist in a room and he's got like a little dish of like grains of sand where each grain of sand is like a piece of understanding and he's like using tweezers to carefully move it from like this pile to that pile of understanding. And someone comes in the door and they're like, "I've invented AI. Here's this spoon and you can transfer understanding much more quickly." But then there's another door and outside there's like endless piles of sand like in every... Like the Sahara Desert of sand, right? And it's like, "Great, I have this spoon now. I could finally like tackle this little problem." But the actual problem is the Sahara Desert. I feel like that's where we're at. Like the spoon is way better than the tweezers, but the problem is the desert size.

CARRIE: Oh, that's such a great metaphor and imagery. Although, I do think, you know, that might hurt some people's hopes about our progress with cancer research. 

OBI: Well, I'm pretty hopeful. I mean, I'm really happy to trade in my tweezers for the spoon. I think that we will make-- we will continue to push the, the ball, the, the boulder up the mountain and much, probably much more quickly. And we have made tremendous progress in terms of cancer treatments, um, you know, extending people's lives and quality of life through this Her-Herculean effort over the last many decades. And if we can accelerate that, it will really matter to patients. But at the same time, we won't be curing cancer in five years 

CARRIE: Maybe each of us have a spoon that we're using every day. And if everyone uses a spoon, we're certainly moving the desert faster than if we were all working with tweezers. 

CARRIE: So given that I just talked about the importance of us all working together to conquer cancer, uh, in our next episode, we're gonna discuss how AI is also transforming the possibilities of collaborative research among cancer research institutions and what that means for the future of data sharing. Thank you for listening to Cracking the Cancer Code, and thanks again to Dr. Obi Griffith for joining me today