YT YouTube2Text

YouTube2Text → Transcripts → AI for Cancer Care: From Cells to Systems | Columbia AI Summit

AI for Cancer Care: From Cells to Systems | Columbia AI Summit — Transcript

by Columbia Data Science Institute · 8,983 words · 865 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:03>>Hi everyone. Welcome. My name is Noemie Elhadad. I'm
0:08chair of the Department of Biomedical Informatics at the
0:11Vagelos College of Physicians and Surgeons or medical School.
0:15I'm also the director of the AI at VPNs initiative that attempts
0:23to promote AI across all of the missions of the school, which
0:27are research, clinical and care delivery, education and
0:31training, and community health. I'm very excited to welcome our
0:37speakers and our moderator for this session on cancer care.
0:41From cells to systems. I think it it fits very nicely. After
0:46the previous session on brain and mental health, because you
0:50will see a lot of the themes come back. And in particular,
0:53this idea that I know is pervasive, from the basic
0:57science and discovery to translation into clinical care
1:01and to shifting from now that we have algorithms that are
1:07powerful enough to be accurate, how do we know? Think about the
1:11connection to humans and go from an algorithmic centric view to a
1:15human centric view of AI, and really serving our patients? I
1:21am going to start the session with a technical talk from
1:25Doctor Raul Rabadan, and then we'll have a panel moderated by
1:30Doctor Gursoy. I want to introduce Doctor Rabadan. It's
1:35my pleasure to do so. Doctor Rabadan and I started as
1:38assistant professors many years ago. I won't say how many years
1:41ago, Raul, but it's been a thrill to see how much Doctor
1:46Rabadan has been able to do. Doctor Rabadan is a professor of
1:50systems biology and biomedical informatics at VPNs. His
1:53research is in computational genomics, and he leads the
1:59program for mathematical genomics. At VPNs, he's at the
2:02forefront of innovation on multiple ways in which AI can be
2:07used in computational genomics, and in particular, in cancer. I
2:11want to do want to come to the welcome. Thank you. Thank you
2:21very much, Noemi. We are in a very exciting moment, as you
2:24probably know. I mean, it's research has been changing, but
2:28this society as a whole, it's been changing through this AI
2:31revolution. I'm going to be talking about some of the work
2:34that we are doing here at Columbia by myself, but also
2:38mentioning some of the work in the program by my colleagues
2:41here is very interesting with applications in cancer, but is
2:44many more things. I mean, with this idea of foundation models,
2:48we can create getting a lot of data like language models we can
2:51create, get a lot of data, train these models in this amount of
2:54data, and then trying to different applications, the
2:57applications that I'm more interested in because I'm
2:59working on this domain is cancer, but it's across many
3:02different biological disciplines. And one of the
3:05things that is happening that is quite interesting is biology has
3:08been always a science that is very descriptive. I mean, we get
3:11data from a patient, we can get single cell, we can get the
3:14genome, we can do clinical records and we see what what are
3:18the things there. And then based on this evidence we try to
3:20decide what is the treatment or what's happening or what is the
3:23diagnosis. But one of the things that is changing is now we can
3:28start doing predictions. My background is in theoretical
3:30physics and in physics we have very good models for
3:32predictions. Once you have good models of how things work and
3:35you can do things, you can send missions to Pluto and arrive ten
3:38years later and they arrive there. They are not doing
3:40somewhere else. So I think here we can do something very
3:44powerful. I mean, once we have very powerful models, the first
3:46thing that we need for these models and these things is data.
3:50And as you probably know, we in the beginning of the century, we
3:54started the human genome era. So the technology was there and the
3:58infrastructure was there. And then many people started
4:00sequencing the human genome and then human genome because
4:04cheaper and cheaper, cheaper. And then now we are sequencing
4:06many, many, many patients in cancer in particular. And people
4:10have been collecting here and in many other institutions, a lot
4:14of data from cancer genomes. So you probably know cancers, they
4:18are caused by mutations, mutations that we inherit from
4:20our parents. They are the germline mutations. These are
4:24contributing to risk if you can. Certain mutations can give you
4:28more risk of certain cancers. But most of the mutations that
4:30we are studying cancers, they are the somatic mutations, the
4:33ones that we get in our lifetime. If you smoke, you go
4:36to the sun, you inhale certain substances. All these things
4:40cause mutations and you age also, they are most of the
4:44mutations. They are related with age. So many people across many
4:46different institutions here in the US and across the world,
4:49they have been sequencing patients. And we have now
4:52probably close to a million patients that have been already
4:54sequenced. And then people are doing these cartographies. I
4:56mean, getting all these. You have the patient, you look at
5:00the mutations. These are somatic mutations in a patient, in
5:02another patient, in another patient and say for one few
5:06cases in pediatric cancer is almost impossible to find
5:10two patients with the same set of mutations, they are
5:12all kind of different sets. But there is a structure in
5:16this data. There is a lot of structure in this data. So
5:18now what we have this cartography, these maps of
5:20how these mutations happen. We have to interpret and to
5:22make sense to this. All this data and correlate with the
5:26prognosis, response to therapies and many other
5:28things that happened. One of the things that have been
5:30very interesting is in the dynamics. I mean, cancer is
5:32our own cells getting mutations. And these are
5:34dynamic processes, mutations, they occur at
5:38some point the cells start proliferating. They are
5:40different clones, different therapies select different
5:42clones. And these are very dynamic processes and
5:46longitudinal studies. They are very informative of what
5:48is important and what is not. But one of the crucial
5:50challenges here is that we don't understand very well
5:54what most of these mutations are doing, and what is the
5:58biological data that is there. So what we have now
6:00in biology, we have very, very good technology. We
6:02have high precision observations with mutations.
6:04We can say for mutation is there or not. What is the
6:08fraction of cells that have these mutations. We can have
6:10we can identify using traditional machine learning
6:14patterns in this data. We can some descriptions. And
6:16we if you see for example single cell data, you see
6:18all these maps that happen with single cells and all
6:20these things. And we can start doing models. Okay. We
6:24don't know what is the final methods that they are going
6:26to be kind of illuminating or predicting. What is the
6:28best response in in my background in physics, for
6:32example, when Newton discovered kind of the
6:34Newton laws, he has to develop calculus and it was
6:36taken like a century to make calculus kind of be a solid
6:40mathematical tool that can be used for solving
6:42differential equations or predicting different things.
6:44Here with AI, we are kind of looking at what is the right
6:48model. And what is fascinating about the field
6:50is that we have many different people coming with
6:52different models. Foundation models is one of those and
6:54is probably going to be the beginning, but we don't know
6:56if this is going to be the final model or not. From the
7:00theoretical point of view, we don't know very well how
7:02how these foundation models or in general deep learning
7:04works. I mean, why is learning, why these
7:06conditions, why these structures? We don't know
7:08many of these things. So they are theoretical
7:10innovations that are coming along the way. I mean,
7:12better models, cheaper models and more, more
7:14structured models. And the idea is to find fundamental
7:18principles that can organize all this information. So
7:20coming from the very beginning, from generating
7:22high throughput data, we are there extracting some of the
7:24data. We are just generating some of these models and
7:26then doing predictions for the future. And we have
7:30several models here. Some of them work here at Columbia
7:34by my colleague here, Mohamed Al Qureshi, looking
7:36at the structures in different proteins,
7:38predicting very much like AlphaFold, predicting
7:42mutations, specific mutations. What are they
7:44doing in protein structures of some of the work that we
7:46are doing and looking at noncoding mutations in
7:48cancers? I think one of the key ideas is this idea of
7:52generalizability. We have a lot of data. We can look at
7:54patterns in data, but we can extract some knowledge from
7:58these patterns and go into something else, something a
8:00patient that you never seen before, some structures that
8:02you never seen before, the applications in cancer and
8:06beyond their manifold. I mean they are from DNA
8:08predicting things. Some of the work that I was
8:10mentioning by Mohammed and myself, and both from
8:14academics and from industry, there is this crazy
8:16competition between different companies and
8:18different institutions trying to create better
8:20models, trying to predict things from DNA, proteins,
8:22the structure of RNA and single cell from molecules
8:26and chemicals, from and to clinical data and
8:30pathological pathology data. This is doing very good
8:32models. This is one of the examples that was the Nobel
8:34Prize last year, that you can predict accurately the
8:38structures of proteins, and then you can start doing
8:40perturbations on all these structures and trying to see
8:42how these mutations that we find in cancer can be
8:46affecting that up. So here. So in the last few years
8:48with colleagues here at Columbia and beyond, we've
8:52been mapping, doing sequencing many of these
8:54tumors and trying to understand what is the
8:56structure of all these tumors. I mean, some of the
8:58work that we've been doing in brain cancers, in
9:02leukemias and lymphomas and all this information, now we
9:04are structuring, structuring and generating these
9:08foundation models based on a lot of single cell data,
9:10trying to see how the different transcription
9:12factors in different specific cell types, they
9:14are kind of associated to different programs in
9:16different cell types. And then we can induce we can
9:20introduce the mutations that we found in patients in
9:22silico. We can try to see what is the experiment that
9:26doing experiments in silico to probe what is the biology
9:28of these things. We have a whole program here at
9:30Columbia where we are just generating these foundation
9:34models and then applying it to many different things
9:36brain tumors, DNA repair genes, pan cancer studies
9:40and MDS, AML, stem cells, prostate cancers and
9:42lymphomas where they are different groups, both
9:46experimental clinical groups working on trying to
9:48understand what is the biology that is underlying
9:50these diseases, using these foundation models. Summary.
9:54We are in a very, very exciting moment where in
9:58cancer research and anything else in the clinical side,
10:01large data is enabling new models. The structure is
10:04there. We can do beyond descriptive to do something
10:08that is predictive. And models are very
10:10quantitative. When we are predicting the effect of
10:12expression, it's not it's not only expression is going
10:14up or down, we can predict exact values of things that
10:18are going there. This is in cancer research. There are
10:21tons of opportunities here. I mean we can generate these
10:24models, can guide experimentation, including
10:26experiments in silico. We look at the patient, we take
10:30the data, and we introduce that in our system and try
10:32to see what's what's going on is allowing scaling and
10:36throughput beyond current limits. At the moment. If
10:38you want to do Crispr, you have to do one mutation,
10:40another mutation. Here we can introduce some of the
10:42mutations simultaneously in the genomes. It helps to
10:46integrate and collect and connect and data modalities.
10:49I mean, many of us we are collecting and working in
10:52many different aspects of many diseases. And what we
10:54are learning is just collecting all this
10:56information. We can learn from some of the data in one
10:58part of our institution to about something that is
11:02happening in a different part of the institution and
11:04all around the world. And I think the next phase also is
11:07quite interesting, is going to the generative part that
11:10is designing new promoters, genes and proteins that can
11:12be doing specific things in a specific patients. What we
11:16need, and I think Colombia is poised to have these
11:18things. We need multidisciplinary teams of
11:21quantitative and experimental teams and
11:22clinical teams, all of us kind of addressing common
11:26problems with different modalities and different
11:28ways of thinking. And the other thing as an academic
11:32institution is probably this revolution is not coming
11:34from us of some of our colleagues here. We are
11:36going to be kind of the older people. It's going to
11:39be the new generation of researchers that are coming
11:42with new ideas, new methods. And our role here is to
11:44guide them to form and guide all these new generation of
11:47scientists as an academic institution, a research
11:50institution, I think we are kind of the best to do this
11:52kind of work. That is going to be the leading work in
11:56cancer research and clinical care in the future. Thank
11:58you very much. And this is part of my team here. Thank you
12:08so much, Doctor Rabadan. I just want to highlight the fact that
12:10something that Doctor Jeannette Wing said when she started the
12:14session, which she's the short time it takes for AI fundamental
12:20research to make it into other fields. Right. And so if we
12:24think about the first foundation models and large language models
12:27that came from text, we can be purist and say it's starting in
12:322017, 2023 is when the world is like, oh, there's such a thing
12:36as a large language model, and now we're in 2025, and we have
12:40researchers like Raul who are able to create foundation models
12:44of the language of biology, which is highly innovative. So
12:48thank you for that. So we're going to move on to the panel
12:52part of this session. And it's my pleasure to introduce Doctor
12:56Gursoy who is Assistant professor of Biomedical
12:58informatics at VPNs and at the New York Genome Center. I would
13:12like to invite all our panelists.
13:34Thank you, Noemi, and thank you all for the great presentation.
13:36And before I start shooting my questions, I'm going to first
13:40very briefly introduce our panelists. I'm not going to do
13:46justice because we have really highly accomplished researchers
13:48here, but I'm going to very quickly do that. And to my left,
13:52Doctor Chin Hur is a professor o medicine and professor of
13:56epidemiology, a recognized expert in cancer screening and
14:00prevention for gastrointestinal cancers, and is a leader within
14:04the National Cancer Institute consortium of researchers that
14:07work towards improving cancer management. And to his left is
14:12Doctor Rita Kukafka, who is my colleague at Biomedical
14:15Informatics, who is a professor and also with a joint
14:18appointment in the Department of Social and Medical Sciences at
14:20the Mailman School of Public Health. And she is leading a
14:24research program at the intersection of population
14:27health, data science, predictive analytics, decision science,
14:30behavioral science and genomics. And I'm going to skip roll. You
14:36already heard. So the next is Doctor Despina Kontos who is a
14:40professor of radiological sciences and of biomedical
14:44informatics with expertise in AI, machine learning and big
14:47data analytics for multimodal imaging data. And last but not
14:52the least, Doctor Anil Rustgi is Herbert and Florence Irving
14:57director of the Herbert Irving Comprehensive Cancer Center at
15:00New York Presbyterian and Columbia University Irving
15:04Medical Center, and a world renowned leader in the field of
15:08gastrointestinal oncology. And all of our busses here. I guess
15:12we are all part of cancer center. So with that, I'm going
15:18to actually start my question to Doctor Rustgi. As the director o
15:22the Herbert Irving Comprehensive Cancer Center, so could you give
15:26us an overview of the main areas of research and patient care at
15:32our cancer center in general and in cancer care, as well as
15:36within those areas? What role the data and AI currently play?
15:40Sure. Thank you. By way of a brief background. Our cancer
15:44center is a matrix cancer center that extends across Columbia
15:49University and involving 7 or 8 schools. Although the crux of
15:54the membership and participation is at the medical center,
15:58there's over 200 members, 400 trainee members, and we're
16:02divided into four programs that basically span the continuum
16:06from very fundamental discovery science and cancer genomics and
16:12Epigenomics and Doctor Robertson's co-leader of that
16:14program, and working with Doctor Al Qureshi on, as you heard, on
16:18computational protein structure and function. So that's a point
16:21of emphasis in our cancer center and emphasis as well. And then
16:26there's another program, precision Oncology and systems
16:30Biology, that's co-led in part by Doctor Peter Sims of Systems
16:34Biology there. The tools of AI are being more employed in terms
16:39of investigator initiated clinical trials. So coaxing
16:42discovery science in the pre-clinical domain into the
16:48possibility of application and phase one clinical trials. So we
16:53put a lot of emphasis on investigator initiated trials.
16:55And we're seeing enhancement of use of AI in that context. And
17:00then I would say we have a clinical informatics unit that
17:08involves Doctor Cantos. So she'll speak much more about
17:10that. That really involves the use of AI in two large areas.
17:15One is imaging led by her in the cancer center. And again, she'll
17:21elaborate, but the other is in fundamental aspects of clinical
17:24care and electronic health records. Can we reach a state
17:28where by if you're the provider and the patient is front of you
17:32with various diseases or cancer states, can you bring up
17:38algorithms for management? What we also call molecular matching.
17:41If they have a certain mutational profile, can we match
17:45them with AI tools to the available clinical trials. So
17:50really it's incorporating AI from discovery science to
17:54translational medicine to fundamental aspects of clinical
17:58care. And finally we have a program in cancer population
18:03science. Can we then take these applications of AI into the to
18:08the our population writ large and have an impact on cancer
18:14screening and cancer surveillance before we get to
18:17the stage of cancer therapy? Thank you so much for this great
18:21overview. And now we can actually dive into the
18:24specifics, because I think we have representation from all
18:26that you touched on. And I want to start with Doctor Hau here. I
18:30was wondering if you could share how you use AI in your work,
18:33especially for cancer screening and prevention for
18:36gastrointestinal cancer? What specific research questions that
18:41AI is helping you answer, and what new possibilities that you
18:43see in the future? Yeah, so cancer is very difficult. I
18:49think in some ways it's in some ways the most difficult disease
18:53area. And so hopefully I will have the most to contribute.
18:55That's the that's the big picture. So I'm a practicing
18:58gastroenterologist and a physician scientist, but I do
19:01mainly cancer population. So I started from an epi type
19:06regression modeling type of machine learning. But then as my
19:10bread and butter was, we built models, simulation models of
19:15cancer in the US population. So we would calibrate. So I
19:19developed a lot of understanding of models. Not quite as smart as
19:22Raoul, but we do something similar. We build a natural
19:26history models of cancers and reproduce the US population and
19:29then test different strategies, including screening strategies.
19:33What age should you start, how frequently for the different
19:35types of cancers? So I've done that in esophagus, stomach,
19:39pancreas, colon and liver. Those are the GI cancers. So the goal
19:43the goal is early detection and prevention right. As as Doctor
19:48Rusk said. So how can how can I help in that? Right now we have
19:54we have risk prediction models. Some of it there's a there's a
19:59difference between early detection. You find the breast
20:03cancer lump earlier versus metastatic. So that's that's
20:07important because there's survival benefit. But then in a
20:09lot of GI cancers for example colon cancer we want to detect
20:14the colonic polyp the precursor before it turns into cancer. And
20:17then therefore it's prevention. So I'm a little biased. Colon
20:21colonoscopy is a way to go. So that's a little different. Like
20:25trying to train a prediction model to detect earlier versus
20:31someone that will progress. And the latter is more difficult.
20:35You need observational data. You need time series data. And then
20:38that's where perhaps unknown risk factors. So most of you are
20:43probably familiar with. You can do a multiple a multiple variate
20:48regression using a logistic regression to see who progresses
20:51versus not. But we want to find hidden or unknown risk factors.
20:56And some of it ends up being the black box. But that's that's the
21:00immediate application. So anyway, that's that's the quick
21:05question. I work with someone I don't think Cheetham Park is
21:10with me, but she's the real data scientist. I'm just the crazy
21:15MD. But thank you. So speaking of screening and prevention,
21:22early detection and prevention. So I want to ask Doctor Cantos
21:25how would we do this from imaging and the modern AI. How
21:29does it help you to do what you do. Great question. And I'll
21:33draw some examples from my own research. And then I can
21:37extrapolate a little bit further. So I'm a computer
21:39scientist by training and have been doing AI almost 20 years
21:44now. Way before it was such a hot thing I guess, and been
21:47living in a radiology department my entire academic career
21:51because I love working with imaging data. So in the
21:55screening setting, for example, there's a lot of information in
21:57imaging data that is not really utilized by the current clinical
22:02evaluation. I'll give an example. I see many women in the
22:05room. So when you go and get your annual mammogram, the
22:07purpose of getting your mammogram has been so far mostly
22:11to tell if there is a cancer there or not, to detect research
22:17for more than 30 years now has been showing that mammographic
22:21density. And you may have heard this in the news as well,
22:23because there's been a lot of discussion about it, is a risk
22:27factor for breast cancer. So the dense tissue in the breast,
22:29which is the fibrous glandular tissue in the breast, as shown
22:33in conventional x ray mammography, can predict which
22:37women are more likely to get breast cancer. And it can also
22:41tell us for which women mammography doesn't work very
22:45well because there their breast is so complex in terms of the
22:49appearance of the tissue. That is very easy to hide for cancers
22:55to hide underneath the complexity of the tissue, and
22:57maybe different modalities that may be more expensive or time
23:01coming, such as breast MRI, need to be to be used. There's a lot
23:07of discussion in the field right now about breast screening, how
23:09it should be used, what frequency, which women should be
23:11getting conventional mammography, which women should
23:13add MRI, which women should get ultrasound, and what age should
23:17we start? No clear guidelines really. If you look at it. And
23:21there's a lot of implications, both about finding cancers
23:25earlier versus later that are much more treatable and also
23:29cost effectiveness implications, you know, and MRIs, older orders
23:35of magnitude more expensive than a conventional X-ray. So we can
23:39give it to every woman. It's not also available as, as broadly
23:43as, as conventional x ray. And so there's all these kind of
23:47complex decisions that need to be made. And this is a perfect
23:51example where AI can be used. For example, in this area of my
23:55own research for many years now, in different in different
23:57settings, how to measure for example, this kind of breast
24:01density, the amount of the dense tissue. If you see how that is
24:05done clinically is literally by visual assessment. And there are
24:07many, many studies that show that if you put like for
24:11radiologists, they're going to assess your breast density
24:15differently. So whether or not there's going to be a decision
24:17that you are high risk woman for breast cancer and you need to
24:21get MRI may not really depend on your actual amount of tissue,
24:23but who's reading the mammogram. So there's a lot of subjectivity
24:27there. So I can take that out of the equation. We can quantify we
24:29can measure it very accurately. We can run the same mammogram
24:33100 times, get the exact same value. So then we have
24:35quantitative values. That's great. But we can also use AI to
24:39combine different types of data that relate to the risk of
24:41developing breast cancer. So we can have more comprehensive
24:45assessment of risk and more accurate. So we can put density,
24:47we can put genetics right. We can put electronic health record
24:51data, we can put family history, we can put all kinds of
24:55information. And AI is really good at looking at the data and
24:57extracting patterns that you cannot see just by by just
25:03conventional assessment or by human vision. And also
25:05mathematically combining the different sources of data in an
25:09optimal way. And so now we can start thinking about really
25:15personalized screening and precision screening. And AI can
25:17help us derive an algorithm that can take with, with, with good
25:19accuracy, which women are at high risk of developing breast
25:21cancer, which women need to get more intense screening, more
25:25frequent screening with additional modalities, but also
25:27look at the other spectrum, which women may be at lower risk
25:31of breast cancer. Maybe that we also be thinking about
25:33intensifying screening for some women because x ray, you know,
25:36does may have, you know, we have the radiation dose there to
25:39consider. It's not as harmful really as one mammogram. You get
25:41the same radiation as flying to East coast to West Coast. So
25:45it's not you know, it's not that much, but still something to
25:47consider. So that's a very practical example of what we can
25:51do with AI. And in terms of cancer screening, for example,
25:53which I think is a great area of application, we can think about
25:57lung screening the same way or other areas that we can leverage
26:01these technology. Yeah. Thank you so much. So while we are at
26:05the breast cancer, I want to zoom out a little bit and ask a
26:11question to Doctor Kafka, because your research is on how
26:13technology can help patients and providers on their own and in
26:17their work together. And a lot of your expertise is in
26:19developing these technologies in cancer prevention. So as we
26:23develop better and better AI systems for cancer risk
26:27prediction, early detection and care, we as a scientific
26:29community are finding out that it is not the end of the road.
26:33Now, the providers and the patients need to handle this
26:37information and act upon it. And in the last session, we touched
26:39upon this a little bit. Can you describe the research challenges
26:43your work tackles and maybe some of the ethical questions that
26:45comes out of it? Sure. It's fun now. Now it's on. Thank you very
26:53much for that question. And so my work is basically leverages
27:01AI standards for interoperability and a patient
27:05centered design to support patients in personalized
27:09decision making and fostering meaningful communication between
27:15patients and their providers. We use it's a data driven approach.
27:23We use algorithms. So far, we've incorporated algorithms for
27:29breast cancer risk prediction into tools. And I have to say
27:33that we've been working on developing these tools probably
27:37for close to a decade, and have run several randomized
27:43controlled trials to evaluate outcomes of how these tools,
27:49when implemented into clinical care, actually improves both
27:53patient reported outcomes and clinical outcomes. I think the
27:59high level umbrella for thinking about the work that we do in, in
28:05my lab is in the space of precision prevention. So we
28:11really try to accumulate the data that we need to run these
28:15prediction models. And we move from a one size fits all
28:23guidelines implementation. For example, for breast cancer
28:26yearly mammography to a much more personalized, patient
28:33centered approach to delivering risk stratified or risk aligned
28:39strategies to reduce risk. And the tools support both patients
28:45and providers. On the patient side, it is decision support as
28:51compared to education because we spend time enabling patients to
28:55deliberate decisions that they have. We've done RCTs around
29:01genetic testing. Chemoprevention. We just
29:03received two additional grants where we're going to incorporate
29:07both cardiovascular disease and breast cancer. Two leading
29:11causes of death in women simultaneously. To see how that
29:16impacts decisions, I do want to focus on a couple of the
29:21components that I think are important and actually feed back
29:25to the to the panel, to the previous panel. And part of that
29:29is the black box. So what we do is the data we use to run these
29:35models include both data from electronic health records, but
29:39we also integrate patient derived data. And the reason, as
29:43many of you know, is that electronic health record data is
29:47often missing data. There's sometimes errors in that data.
29:53So prior to this, two purposes for integrating the patient data
29:56and the EHR data, one is for more complete data set. But the
30:01other part is that patients are able to download their data
30:05using someone mentioned the century 21 Care Act, which
30:09requires that patients can access. So through through our
30:13tool, which is called real risk, they access their own EHR data.
30:17They look at that data, they correct the data as needed. They
30:21add missing data, for example family history, ethnicity might
30:25be missing. And then they actually see the data that goes
30:29into the risk prediction models. Once we. And that requires both
30:35the algorithms and also the standards for interoperability.
30:39Someone also mentioned at the last panel that communicating
30:43that risk is important. We want patients to make informed
30:48decisions. So we spend a lot of time really thinking about and
30:53evaluating different ways to communicate risk. And all of
30:57this, I should say, is outside of the clinical encounter. So
31:01one of the things that we do in this decision is, is they have
31:06interactive games where we rig pictographs so they could see
31:10what a 12% or a 20% risk looks like. So we do spend time we do
31:16the same kind of gamification to talk about risk and benefits of
31:21different risk aligned preventive options. So they
31:25understand their risk. They understand what they can do to
31:29reduce that risk or impact their risk. And then they're
31:34encouraged to take, you know, they have an action plan.
31:37They're encouraged to take that to their provider to discuss
31:40during an encounter. Providers also have decision support,
31:44which I won't talk too much about, but the data is
31:47communicated both to patients and providers. So that's pretty
31:53much and you know, summarized in five minutes, but it's been 10
31:57or 15 years of work and a lot of funding to get here. But I think
32:03that the part that that I think is most unique is that we have
32:09figured out how to leverage AI interoperability standards for
32:14data integration. So we do are able to collect the data not
32:17just from the clinical data, but also patient data that they can
32:21best contribute. And we promote this through a patient centered
32:27design. And I do think that that will bring us to the next phase
32:33of how do we actually move AI into action so that it improves
32:41so that patients can actually look at these predictive models
32:45and, and it becomes actionable both for the patient and for the
32:49provider. Thank you so much. And now while we are at AI in
32:55action, I want to now zoom in to the cells and ask a question. So
33:01I'm sure there are a lot of cancer biologists who got very
33:05excited reading your paper, but they probably don't know how
33:09that can help to their research. So can you tell us what they can
33:13do with these foundation models? Can they design the next
33:16experiment or hypothesis, or what kind of downstream tasks
33:19that people who are not developing the models, but using
33:22the models can, can, can do? Yeah, there are many, many
33:27different directions. One of the things that is quite interesting
33:30being here at Columbia is that we have very interdisciplinary
33:33teams of people. So for example, I can mention we are working
33:37with some groups working in B-cell lymphomas and T-cell
33:40lymphomas, different groups. They have their own interests.
33:43Some of those they're interested in, for example, trying to
33:46understand what is the impact of some mutations in non-coding
33:48parts of the genome. They publish some work, but they
33:52already they're mapping some of these mutations. Most of those
33:55we don't know what they are doing. So we are co-developing
33:58some of these methods with one application in mind. The models.
34:02They are kind of huge. I mean, so you can learn many things
34:05with having some application in mind. Also guide the development
34:08of these of these models. So we are working with these teams,
34:11optimizing the models, applying it, developing some of the
34:14statistical framework for thinking about the impact of
34:16some of these mutations, other groups, they are working in the
34:20microenvironment, for example, trying to see how the other
34:22cells in the tumor, they are communicating with cancer cells
34:25promoting or just the immune system, how the immune system is
34:29working with them. So all these different applications, they are
34:32different ways that these models can, can, can evolve. It's very
34:35hard to develop a model that works for everything at the
34:38moment, but they are kind of working with different teams. We
34:41have different applications in mind, and that's the way that we
34:46are at the moment. We are just following, just working with our
34:48colleagues here, looking for funding, looking for data and
34:53generating data and designing experiments all together. So I
34:55think the collective endeavor here is quite important. Thank
34:59you so much. So now I'm going to ask open ended questions. And it
35:04might be a little annoying for the panelists. But first I want
35:07to ask to our clinician scientists. And I want to ask,
35:13where do you see AI in ten years in terms of clinical care and
35:17feeding? Back to the research that you're doing, doctor Ruskin
35:21and Doctor Who did you doctor would you like to okay. So
35:25that's going to be that's really difficult to answer. So but I'll
35:29try. So a lot of a lot of what I started doing a while ago in
35:34terms of risk prediction, risk stratification that used deep
35:39learning or neural networks. Right. So I was trying to
35:44identify hidden risk factors or things that were not obvious.
35:48But you know, as as we all know, ChatGPT etc, the, the generative
35:54AI and it's interesting. How are we using it now in the, in the
35:58work that I do. So for example, they're very good at cleaning
36:02data because especially if you're doing EHR data, that's
36:05that's the hardest part. When I first came when I first hired my
36:09person, I said, I want you to download all the Columbia
36:14pancreatic cancer cases and just start cleaning because we're
36:18going to mine the heck out of it, right? We're going to do
36:20keyword searches and then correlate it with lab data, time
36:23series data, and predict someone who's going to develop
36:26pancreatic cancer. You know, two, four years later, we're
36:28still we're still trying to clean the data. It's very and
36:32it's, you know, there's thousands and cases. So that's
36:35going to be big. I know that sounds like low hanging fruit
36:38cleaning cleaning data I mean, genomics and radiomics are a
36:41little better because ultimately they're ones and zeros. So we
36:44homed in on laboratory data for example. So I think that's going
36:48to that's going to be big. We have used it also in the Gans
36:51Generative adversarial Network. We've created a phantom images,
36:55particularly when there weren't enough images. But it's a little
36:58tricky when you don't have enough data, you don't have
37:02enough data. So we're trying to I think in engineering this is
37:06used more frequently. But we're trying to create synthetic data,
37:09particularly of patient phenotypes that aren't as
37:13common. And then we can run simulated clinical trials. I
37:18focus on screening, but that could also be applied for
37:22treatment in terms of decision support. Everyone knows like you
37:26know you're in your epic thing and beep, beep, something pops
37:29up. This person's going to develop pancreatic cancer. Order
37:31an MRI now. Yeah, I think I think so I think a lot of
37:36people, that's what they think of digital health. I'm not as
37:40much in that space. I think, you know, a lot of times when the
37:46current risk factors we have are sometimes pretty good. So it's
37:50more about pushing the boundary, finding it earlier. And we need
37:54to be careful. I mean, ultimately, you know, like
37:57computational people, if you have enough cases eventually
38:00central limit theorem, you'll be able to predict everything
38:02perfectly right. But it's going to be a long time, especially
38:06with cancer. It's so complicated. You have genomic,
38:08you have genes that are you're born with. But then
38:14environmental exposures are very, very complicated. And it's
38:18not like asthma, high pollen count. You have a flare. These
38:22these environmental exposures could have been 4060 years ago.
38:26This is going to take a very long time. So I think we need to
38:30at least start until until we have perfect knowledge. I think
38:34we need to start with the low hanging fruit. Like for example,
38:36the example I'll give is I'm quite interested in liquid
38:40biopsy for cancer, blood tests for cancer detection. It's a
38:44very hot topic and we need to be careful. There's some. The first
38:48commercial product screens for 50 cancer simultaneously. And we
38:50need to be careful because every cancer is different and we can
38:54do a lot of harm. When you say someone's positive and you're
38:58not sure if it's a false positive and all the over all
39:02the difficult situation it creates. So we need to be very
39:06careful. For example, going back to colon cancer, if you have a
39:08positive test and it's a false positive, we do a colonoscopy
39:12and then we can assure the patient that they do not have
39:16colon cancer. I mean, nothing's 100%, but we can look them in
39:18the eye and say, you don't have colon cancer, whereas in other
39:22cancers it's so we need to be specific. Like just like one
39:28you've seen one cancer, you've seen one. You need to be careful
39:30about the clinical applications per cancer. I know I know, rolls
39:34going the other way with foundational models and that's
39:36more for biology. But yeah, so that's the long answer. I think
39:42a theme that you're hearing on a recurrent basis is cancer
39:46prevention through screening. And there many meritorious
39:52aspects of that pursuit right now, health care in the United
39:57States comprise constitutes about 24%, 25% of the total GNP.
40:03And yet it's not regarded as the absolute best health care system
40:08in the world. Our mortality, maternal mortality rates are
40:12high. Our our our fetal rates are high. All sorts of problems
40:20with aging. And I think the a bold vision requires
40:26intervention using AI tools as it is increasingly in the
40:30consumer world, in the health care world. And that will lead
40:34to an impact that will be positive on the national
40:40economy. So I think we've got to marry AI with very practical
40:44health care considerations at a societal level. Thank you so
40:50much. So now we're talking about the ten years vision of AI in
40:56cancer. I want to ask Doctor Robert a question. So ten years
40:58ago when I was a young grad student, when they they told me
41:02that, oh, let's do protein structure prediction. And I
41:06could never imagine at the time that AlphaFold is going to be a
41:10thing. And I said, oh, this is such a boring problem. Everyone
41:12is working on it and I don't want to work on it. So then I
41:16ended up working at Encode consortium, where we were trying
41:18to basically functionally annotate all the non-coding
41:21mutations. And again, I saw that as maybe not a problem that we
41:26can solve. So do you think that in ten years, with the help of
41:28AI, are we going to understand the functional role of every
41:32non-coding mutation in our genome? I don't know, but I
41:36mean, I think we are getting closer to that. I think it's
41:40clearly I mean, there is a lot of information in coding in,
41:42let's say, for example, something that is happening in
41:46cell, I mean, all cells, they have most of the cells, they
41:48have the same DNA, they express different genes. And there is
41:51information in every cell and different cells. They have
41:54different parts of information that is informing other ones,
41:56other organisms, mouse and humans. I mean, from the point
42:00of view of cells, they are not so different. I mean, many
42:02things are. So we can leverage some of this information. There
42:06are many models, organisms, even a fly or C elegans. I mean some
42:08of the biology is the same. So if we can start incorporating
42:12all this information, we can get very rich annotation and
42:16functional annotation of some of the role of some non-coding
42:20parts in the genome. So I think in the future we will get
42:22closer. I mean, that's for sure. I mean, how close to the final
42:26response? We don't know. But I think closer and as we are
42:28getting richer and richer, I mean, the model that I was
42:32presenting is only human, but the team now is working on mice,
42:36creating one for mouse and perhaps just getting more
42:38information about different species is going to enrich our
42:42knowledge of the biology and different conditions. I mean,
42:44cancer is a particular condition, but you said you have
42:46an infection. For example, cells in an infected condition, they
42:49are very different from the other ones. I mean, so once we
42:52get more information, we can start learning some of the
42:54structure that is there and different people and different
42:58communities and different ethnicities, I think is quite
43:00important. So the more diverse information, the more we can
43:04learn about all these parts of the non-coding genomes. Thank
43:08you. And Doctor Cantos, where do you see in ten years, the
43:12radiology and I in the same way. So first of all, to say
43:16something about radiology and I, I didn't even know this, but 80%
43:20of the FDA FDA cleared AI tools are on radiology applications,
43:28which is great. Right now, as a computer scientist who is who
43:32lives in radiology, you know, one of the reasons that I chose
43:36to live on the on the medical side of the of the spectrum is
43:40because I would ultimately like to see some of the tools that we
43:44make in a laboratory setting, make it to the clinic and be
43:48something beyond just a publication or something that
43:54inspires, you know, the development of tools that
43:56actually make it to the clinic. And so I would like to see a
44:02more rigorous process of evaluating the algorithms we
44:06create, the tools we create in terms of putting them in the
44:10clinic and more scientifically kind of established path for
44:14doing that. But at the same time, a more streamlined way of
44:16doing that. So, you know, right now we're not good in either. We
44:20don't have a very rigorous way of evaluating these tools in
44:24general, and not just like the academics, like everybody. And
44:28we also don't have a very well streamlined way, you know, it's
44:32amazing. So when I was a grad student, I decided to go to a
44:36medical school for my career because I saw radiologists doing
44:40the easiest measurements by hand. If you don't know what a
44:44resist measurement is, as long as you are so in in in treatment
44:48for cancer and indication of whether or not a patient is
44:52responding or not is to estimate the longest diameter of the
44:56tumor. And if it's shrinking, is working. If it's not, it's not
44:58working. So the way they were doing that would literally take
45:02the mouse on the screen and click, click on the MRI on the
45:06workstation visually. So if I'm doing it versus you doing it
45:10could be Friedman's working treatment's not working. It's
45:14kind of still happening the same way. And we have I don't even
45:18know how many I don't know thousands of segmentation tumor
45:22segmentation methods published in the literature. Right. I
45:24would love to see that gap close as much as possible. And I do
45:30think this is an area that academia has a strong role to
45:32play in creating more creative partnerships with industry. I am
45:38very passionate about that. To finally see some of these things
45:40make a difference. Thank you so much. And along the same lines,
45:46Doctor Kafka, as we close this gap, how are we going to
45:50interact with the patients about AI in ten years? What is your
45:54vision? So my vision is a prediction that might seem out
46:02of maybe off the wall, but I, I think that there is potential
46:12and I think it will might happen that. A.I. Will be and it's
46:22happening now I mean direct to consumer AI to the point where
46:28as health care organizations become more difficult to
46:34navigate as primary care providers are less and less as
46:40patients have more difficulty accessing health care. And we
46:45see this now that this these, you know, ChatGPT large language
46:50models are. We see startups promoting AI directly to
46:58patients for diagnosis, for treatment. And I do think we're
47:04going to see more and more of that in part. I mean, there's
47:10many logistical reasons. Not the least is the amount of capital
47:14that industry has to test different applications, which
47:20becomes more difficult in academic settings. That we see
47:26that some of these industries incorporate providers so they
47:32could prescribe, they could treat. So I do think that as we.
47:40As our perspective is, is largely from an academic and
47:44health organization perspective, the world is moving and it's
47:50moving outside of healthcare organizations. And I think AI
47:54and access direct to patients is going to be commonplace in ten
48:02years. Can I just add something? I think the reason it's so
48:06difficult because you said ten years and I said, you know,
48:09that's hard. I think it's, as we all know from the press and
48:12whatever the, the generative AI, we it's growing so
48:16exponentially, we don't know what ChatGPT five zero is going
48:20to look like. But it's growing so rapidly and we don't even
48:24know how to conceive of how it may help us, I think. So this is
48:28this is a good thing. I mean, we don't know. I mean, we haven't
48:32even figured out how it might help us, but especially when
48:34they have better inference. Right now they're training.
48:38It's. But, you know, I think the, the, one of the obvious
48:40things is cancer drugs. And this is more space. But I think I
48:45think that's I think it should happen. I can't conceive of
48:49exactly how I don't know if you can. But anyway, I want to, but
48:52this is an exciting time. I think this is a really exciting
48:56time because, you know, I'm used to thinking along deep learning
48:58and we can find more hidden data. But this is this is
49:02another once we get inference, what what is possible, we don't
49:06know. But as a physician, if you were to pick a problem that is
49:10not unsolved and ask computer science community to develop
49:14something, what would be that problem in cancer care? Yeah,
49:16well, I'm a prevention guy, but I think, I think drugs would be
49:22fantastic. You know, if you can, if you can cure cancer and
49:26whether whether it be drugs or genetic modification, whatever.
49:31And that's I think that's where a lot of precision oncology is.
49:33We're trying to understand pathways so that we can design
49:38drugs that are specific to the pathways and mechanisms. But,
49:42you know, these are very complex systems. As Raul is a pretty
49:46smart guy. These are very, very complex systems. And I think
49:49we'll have a lot of elucidation of the, of this in a highly
49:52accelerated rate role. I mean, I'm speaking out of turn. So I
49:56mean, tell me if you disagree with some of what I'm saying.
50:00No. No. Like you agree or no. Like, I'm, I just don't know
50:07what I'm talking about. No, I think we don't understand very
50:10well what is the future. I mean, it's very hard to predict. Yes.
50:14So while I have you all so there are a lot of virtual cell
50:18initiatives now, you know, in the field where people are
50:21trying to build a virtual cell using foundation models. Do you
50:26have any of that kind of work in your lab or what do you think
50:29about those? No, I'm not involved. I know that Muhammad,
50:32the person I was mentioning, he was in one of the papers. I'm
50:36not involved. I'm just looking at only specific some particular
50:40two domains or three domains. So genetics, epigenetics and
50:42transcription at the moment, doing something much more
50:46ambitious I think requires big team. I mean, it's probably
50:49going to be kind of a Manhattan project of many different people
50:51going from modeling protein structures. I mean,
50:54organizations cellular organization is going to be very
50:56complex. It requires kind of a an army of people. I have a much
51:02more modest approach. I want to ask any any last thoughts on the
51:08future of AI from any of you or the AI at Columbia in cancer
51:14research, and what do you want to see, or what are the things
51:19that are exciting you right now? Again, another open ended
51:25question. So something that's exciting and a little scary. I
51:27think foundation models might put a lot of PhD theses out of
51:30out of work. Right. You know, we used to have like theses around
51:36for example, in my field, medical image registration,
51:39medical image segmentation, whatever, and all these things
51:41are going to be simplified. Much easier done. Which I think
51:44leaves us with answering the more interesting questions using
51:49AI. Right? So there's a lot, you know, in AI for doing some basic
51:54tasks. We're going to see a lot of that, which is great. I think
51:58that's going to give us the freedom to think more creatively
52:01and asking interesting questions. Yeah. I want to
52:04mention something that I mentioned in my presentation. I
52:07think we are in a very exciting moment and a very exciting
52:10place. I mean, here we are developing these tools where
52:13people from computer science, we have a great computer science
52:16department, people in math, applied math. We have people
52:19doing basic biology, people doing clinical work. We are in a
52:23big institution with many different expertise. And
52:25developing a tool is not something, just developing a
52:27tool itself. I mean, it's in the context of different
52:30applications. And I think an academic institution with all
52:33these different domain expertise is the right place to be. And as
52:37was mentioned, we don't know if these foundation models that we
52:39have right at the moment, they are going to be the right tools.
52:42This is evolving very fast. We don't understand the foundations
52:45of these things. Why this architecture and not another
52:48architecture, why these parameters and not other
52:50parameters is driven. And this is very exciting by
52:52practitioners. I mean there is the theory is basically absent.
52:55So from people doing some of basic foundation of this
52:58foundation models or these deep neural networks doing something
53:04from the basic for the applications and having a
53:06university with this portfolio of intellect, doing all these
53:10different things. I mean, having the right teams here is what is
53:14going to drive the field. And it's probably not being myself
53:17or others. It's going to be the new generation. So I think it's
53:20very exciting for the new people coming into the field because
53:23they are coming with new ideas, new methods that we are not even
53:27thinking about, and combining the expertise of someone working
53:29in clinical care or someone in imaging or doing what we are
53:32doing or what you are doing, and I think now is kind of a wild
53:36field, and I think it's a very, very exciting moment. It's a
53:39revolution and this is the right moment. It's just in the middle
53:42of the mass of the revolution here. Well, thank you so much,
53:47all of you. And thanks everybody.
53:50[applause]

About this transcript

This page contains the full transcript of AI for Cancer Care: From Cells to Systems | Columbia AI Summit by Columbia Data Science Institute, generated from the public captions YouTube serves with the video. The transcript has 8,983 words across 865 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.