[00:00:00] Adam Walker: From Susan G. Komen, this is Real Pink, a podcast exploring real stories, struggles, and triumphs related to breast cancer. We’re taking the conversation from the doctor’s office to your living room
[00:00:15] This is Real Talk, a podcast conversation where we’re digging deep into breast cancer and the realities patients and survivors face every day. We’re talking openly and honestly about just how difficult breast cancer can be, from being diagnosed to selecting the right treatment plan, to living day to day with metastatic breast cancer and life after treatment ends.
[00:00:35] Artificial intelligence is buzzy. There’s a lot about it in the news, and a lot we think we understand about how to use it, and a lot we may not fully understand, but the promise it holds is tremendous. And today we’re going to learn a little more about how it could benefit breast cancer research and patient care in the near future.
[00:00:54] Doctors Charles Perou at UNC Chapel Hill, and Erika Crosby at Duke University are two very accomplished breast cancer researchers who are joining us today to share their perspectives on using AI now and in the future. Dr. Perou, Dr. Crosby, welcome to the show.
[00:01:10] Dr. Charles Perou: Let me start. I can tell you a little bit about myself, right?
[00:01:13] So I run a breast cancer, I’ll call g- genetics and genomics research laboratory here at UNC. I’ve been doing this for, God, 25 years now. And over those many years, we’ve done a lot of statistical learning, machine learning, and now we’re doing artificial intelligence and all those other things I just described as well.
[00:01:37] AI is coming certainly very much into lots of research use and becoming part of clinical use, and that also has many sort of clinical considerations, ethical considerations. Particularly because for a lot of these clinical uses, we wanna look in and look at everyone’s electronic medical records and extract important new information from there sort of synthesize it in new ways that haven’t been done before, and and then act upon that.
[00:02:21] And so that’s the new way of doing things, and there are significant privacy concerns To to be paid great attention to, right? Because we don’t want that information to get n- get out there, to be used properly and to fulfill the the powerful approach that that artificial intelligence is giving us.
[00:02:53] So you’re on an you’re on an ethics committees- Yes. … overseeing the implementation of new artificial intelligence tools?
[00:03:00] Dr. Erika Crosby: What’s happening is I do all of the curriculum development for our responsible conduct of research courses. I direct all of the courses for our first-year and fourth-year PhD students, and so that spans all of the PhD programs in the School of Medicine.
[00:03:16] And so we have bioinformaticians and statisticians that are, this has been their daily life forever, and then we have the molecular biologists who are like, “Ooh, and now I can write code too,” and so it’s a whole, it’s a whole gamut of having h- the conversation, and students are sitting, some are in the School of Medicine, some are in School of Engineering.
[00:03:43] And right now at Duke, the kind of oversight and technical guidance is different depending on what school you’re in, and it’s kind of changing month by month as new things come out. And so students I end up being a touch point for students to be like, “Am I allowed to do this? Am I allowed to do this?
[00:04:02] Can we– Which ones can we use?” and so I just, I’m trying to stay on top of what the today’s recommendations are from the School of Medicine. Well, I think you, you bring up some good uses there, right? I mean, I mean, maybe for our listeners we can mention some of the many sort of existing uses of AI in breast cancer or just cancer research and then maybe some of the things that are on the horizon.
[00:04:33] Dr. Charles Perou: For example, I do know there is artificial intelligence is now approved for use to sort of augment mammography reading, right? So- Yeah … so there’s new AI tools to help interpret mammographies and to identify the susceptibility to develop breast cancer from the mammography image is just one.
[00:05:00] Another is it’s being increasingly used to do what pathologists do, right? Which is to look at what we call H&E images of tumors that are standard images taken in the pathology lab that are used to make important you know- characterize tumor phenotypes and then oh, it’s a big tumor, oh, it’s an aggressive looking tumor, it has lots of what we call mitotic figures.
[00:05:36] It has lots of dividing cells.
[00:05:38] Dr. Erika Crosby: Yeah.
[00:05:38] Dr. Charles Perou: Right now we don’t need a pathologist to, to do that. The the artificial intelligence can do that. Yeah. Another one that I, that you touched on is large language models, right? This is what the GTPs are and what we often use to help us writing, right?
[00:06:02] And of course, that brings up ethical and schooling considerations, so to speak, right? How can we effectively and honestly use these tools who have the capability of writing many of these things all by themselves? Yeah. I’ll say in the clinic, right, a lo- a lot of these la- large language models can be used for clinical trial matching, right?
[00:06:25] Yeah. So it can look at your records and say, “Oh, based on these features, I think you could be eligible for that trial.”
[00:06:32] Dr. Erika Crosby: They’re using it to search to find appropriate clinical trials for patients, so it’s not relying on the doctor knowing what’s available or the doctor knowing the doctor that’s running the trial.
[00:06:45] It can kind of just look very broadly, not even just at Duke, but more broadly at what tr- clinical trials patients could match to, and I know we’re really excited about that because clinical trial enrollment is, was one of the big hurdles, and I think getting the word out or finding appropriate people for them is definitely a hurdle.
[00:07:04] So I know that’s a big, that’s a big use in the clinic at Duke right now too.
[00:07:10] Dr. Charles Perou: We had a, we had an AI symposium here a few months ago, and a couple of the speakers… Institutions were using it to, to largely, like, scan the medical records and just from the existing records, you could kind of identify patients who were just at risk of having adverse events, right?
[00:07:35] Whether it just be, It had sort of an ability to find the frail patients. Yeah. Just again, just by looking at the records, and I think that should help everyone, right? I mean, I think in many instances it’s obvious to the physicians and the nurses who is frail, right? Yeah. But in a few instances it is not, right?
[00:07:55] Yeah. So AI, like a lot of these tools, does… A lot of what it does you can already do but it does it a bit more automatically and- Yeah … and it finds a few things we all miss, I think that’s where it’s really going to be helpful.
[00:08:12] Dr. Erika Crosby: We also, to search, so to do kind of retrospective studies where we look at hundreds and thousands of patients and try to look for patterns, in the past has been somewhat cumbersome, and the things that we use to search the records at Duke, I don’t know how they are at UNC, but at Duke they were like, it’s like an hou- several hours-long courses, and it’s all Boolean logic, and you could get to the end and end up with zero people because you missed something in the middle.
[00:08:40] And they just rol- they, they rolled out a, it’s essentially like a large language model. It’s called Scout to help you search to do retrospective studies, and it, you can have a conversation with it to pr- fix your parameters, and I think it just removes some of the barriers of having to have the perfect search and the perfect training and these clunky tools, so.
[00:09:06] Dr. Charles Perou: Yeah, that sounds great to me because I used to do it the old way you just described there. I would spend many hours putting in a keyword or two to search Exactly So actually, Erika let me ask you a question, right? So you- Yeah … you run a basic and translational research lab. But what, how, h- how is AI like- Making it into your lab for everyday use aside from some of these- Yeah
[00:09:33] that you mentioned?
[00:09:35] Dr. Erika Crosby: I personally use it a lot for writing, as you said. There are a lot of tasks that it can help with. My students are definitely using, we, we use Claude we like Claude Code to help with fixing code and I’ve seen students in the last three months be able to make more progress than they did in the previous year because they spend so much less time looking for the extra space or the something- Yeah
[00:10:04] that broke their code. They just put it into Claude and Claude says, “Oh this is what’s breaking it. This is how you fix it.” and so definitely with code, but we’re also building… You said clinically pathologists are it’s taking the job of pathologists, but we do similar things in the lab where we analyze tumors or we look at tissue, we look at the adipose tissue and we’re looking for metastases, trying to quantify them, and that has been all done by hand in our lab up until now, and we just used Claude to build a user interface that you can tr- you train images and, like, direct it for the first 10 or so and then let it loose, and it can just run through a whole bunch of images and really speed things up.
[00:10:57] And of course, there’s so much we have to double-check and triple-check and make sure it’s actually doing a good job, but it has– I think it’s going to speed things up quite a bit as far as the analysis goes.
[00:11:09] Dr. Charles Perou: So Claude can do image analysis?
[00:11:12] Dr. Erika Crosby: Yeah. Well, Claude can help you build a program that can do image analysis.
[00:11:16] Dr. Charles Perou: That can be… Okay. Gotcha.
[00:11:17] Dr. Erika Crosby: Yeah. Yeah. So it’s like a user interface program that does images. So I know a lot of the students use it to try to cut their– Like, they’ve written something and it’s too long, and it can help ed- with editing. I remember the first time I used a large language model, it was Copilot.
[00:11:36] I don’t know if you’ve used Copilot, and I-
[00:11:39] Dr. Charles Perou: A little bit. Yeah, I use Copilot. Yeah.
[00:11:41] Dr. Erika Crosby: I thought, “This is going to be great. I have a long list of gene ensemble IDs, and I need to know what genes they go to and I don’t wanna Google them one at a time. So I’m just going to put the whole list in and copy the list in,” and it spit out this beautifully formatted table that had the gene ID and the description, and I thought, “This is amazing.
[00:12:05] I just saved myself hours.” And then I go to look at the first one, and it was completely wrong. Like, it had made a really beautiful table full of completely inaccurate information.
[00:12:17] Dr. Charles Perou: So what do you do with that, right? So AI makes AI makes mistakes Or has hallucinations Nope I don’t know, do you have any what do you do when that happens?
[00:12:31] Dr. Erika Crosby: Yeah, I mean, I think, first of all, I think it’s gotten a lot better. I think a year ago we were in a completely different place, and that’s why it’s made the conversation. That’s why we’re talking about it, because it has gotten a lot better. But I think you can’t you need to preserve your critical thinking ability when using it you need to be able to be the ultimate arbiter of what is true and not. So I think that’s what we’re stressing to students, to use it as a tool, not as a brain.
[00:13:04] Dr. Charles Perou: Yeah, completely agree. I’ll say I kind of use it as it’s giving me advice.
[00:13:11] Dr. Erika Crosby: Yeah.
[00:13:11] Dr. Charles Perou: Right? You don’t, you don’t listen, you don’t act on all the advice that you get, do you?
[00:13:17] Yeah. Right? Or you shouldn’t probably. It has a lot to draw upon particularly for the large language models, right? I mean, it has, those are trained from, like, all text, all all medical records. Now, I will say that’s a challenge, right? So if you look at the the published scientific literature, not every paper is good.
[00:13:37] Dr. Erika Crosby: Yeah.
[00:13:38] Dr. Charles Perou: And some topics are, have less depth or knowledge than others, and I think that’s a limit- that’s always going to be a limitation of, of- Yeah … of these AI models, is they’re only as good as the data that’s underlying them. And in sparse areas they’re going to make, they’re going to sound Assured, but they’re not, because they tend to phrase things positively.
[00:14:12] They’re confident, right?
[00:14:13] Dr. Erika Crosby: I’m reminded, so back when I first started teaching medical students, it was the first year, this was when I was still up at Penn, that they gave the med students iPads. And so it was the first time that while I’m standing there teaching, they can, like, Google things. And I’ve, I learned that I had to change my lectures to incorporate whatever the top line of the Wikipedia page for whatever I was saying out loud said, because they would Google it, Wikipedia would come up, they would read it.
[00:14:47] That was ground truth for them. And so I think we’ve we’ve navigated this before. I think this is at an unprecedented level, but you have the tools to think through and have to decipher what is true and what is not, and what’s useful and what’s not. But I’m really optimistic that we’re going to be able to use it to help streamline processes that use time and resources and be able to make things move more quickly.
[00:15:18] I mean this whole drug development research process is sl- is slow, and so I think patients want, they want solutions now. We need solutions now. And so I think anything that can help us be more efficient is good in my book.
[00:15:36] Dr. Charles Perou: Yeah, actually you just brought to my mind one area where AI has been hugely impactful, and it actually won the Nobel Prize, right, which is in three-dimensional protein structure, right?
[00:15:47] So their AlphaFold and some of these other methods, again, utilizing the advantage of AI, right? They could train on all pro- known protein structures- for which there is thousands, probably tens of thousands. And they’re– and thus it was able to build robust models that can now predict three-dimensional protein structure from a one-dimensional sequence.
[00:16:18] And it– that’s really having an impact on protein modeling, protein small molecule modeling, and I think that is Common practice now in the, in pharmaceutical industry to u- to and academia to utilize these AI-informed protein structure tools, yeah.
[00:16:40] I’ll say it part of it in my laboratory, again, we are using it to do a lot of these analysis of images, right?
[00:16:48] I mean, where it sees things the human eye can’t see. And and it just makes it a more powerful microscope, basically. And as long as you can feed it more and more annotated images, it, these things are just going to get better.
[00:17:11] Adam Walker: Thanks so much for being here today.
[00:17:12] Really appreciated having you. This conversation has been fascinating to listen to. So I guess my last question is, what excites you the most about the future of AI-assisted research?
[00:17:23] Dr. Charles Perou: The future of the research is huge and we have a long ways to go, and it’s just a, it’s a spectacular new tool for me, right?
[00:17:36] It’s going to enable us to go faster and be smarter. Now, I think as Erica pointed out, right, this is not an excuse for us to stop using our brains. It’s actually an excuse for us to, to use our brains more creatively- And more, and more thoroughly, right? And so in the lab, it’s going to impact everything from writing papers to summarizing complex topics, and again, in sort of this microscope analogy, helping us see things we couldn’t see before.
[00:18:12] And as we get even more and more sophisticated tools in the laboratory new technologies spatial transcriptomics, single cell technologies new ways to design small molecules, it’s just going to make things, I, better and faster. But again I still see these as integral tools in the laboratory not the driver.
[00:18:43] The scientist is still the driver, in my opinion.
[00:18:46] Dr. Erika Crosby: I could add exactly what you just said. I think we’re, we generate a lot of these big data sets, and the big data has j- is just getting bigger and bigger. And so I think having better tools to help you figure out how to rigorously analyze it and get as many insights as possible from these big data sets is really exciting.
[00:19:10] I’m also really excited about all of the integrations with the electronic medical record. Chuck we’ve talked about this throughout the whole time- Yeah … but being able to get patients access to better therapies or clinical trials that we might not have flagged them for in the past there’s a limit to how many patients a doctor can see and how many, how much experience each doctor has, and there are a lot of these breast cancer treatment re- regimens that are doctor’s choice, right?
[00:19:42] So there’s no clear right answer, and I don’t think this is going to find the right answer, but it’s another tool to help, I think our, support our doctors’ decision-making and get patients where they need to be.
[00:19:54] Adam Walker: Thank you so much for joining us on the show today. This is a really important topic.
[00:19:58] I think it’s important for us to understand the implications of it, and I r- I really appreciate everything you brought to the table Thanks for listening to Real Pink, a weekly podcast by Susan G. Komen. For more episodes, visit realpink.komen.org, and for more on breast cancer, visit komen.org. Make sure to check out @susangkomen on social media.
[00:20:18] I’m your host, Adam. You can find me on Twitter @ajwalker or on my blog, adamjwalker.com.