Early Signals at ASCO 2026 | AI, Multimodal Data and the Next Era of Precision Oncology

What if artificial intelligence could help remove the uncertainty from cancer treatment decisions?

Precision oncology has already transformed cancer care by matching therapies to actionable biomarkers, but today's advances are moving beyond genomics alone. By integrating molecular, clinical and pathology data at scale, AI is helping researchers and clinicians better predict treatment response, identify the right patients for therapies and accelerate the development of more personalized cancer care.

In this episode of The Syneos Health Podcast: Early Signals, Wael Harb, MD, Head of R&D and Scientific Strategy in Oncology at Syneos Health, sits down with Ezra Cohen, MD, Chief Medical Officer at Tempus AI, to discuss how multimodal data and artificial intelligence are reshaping precision oncology. Together, they explore the evolution of biomarker discovery, the growing role of AI in clinical decision-making and what it will take to bring truly individualized cancer care into routine practice.

The conversation also explores how digital pathology, real-world data and AI-enabled clinical workflows are improving patient selection, supporting guideline-directed care and laying the foundation for earlier detection and more precise treatment strategies.

Topics include:

• Why precision oncology extends beyond targeted therapies to reducing uncertainty throughout the patient journey 
• How multimodal data and AI are improving biomarker discovery and treatment selection 
• The promise of digital pathology and AI-powered clinical decision support 
• Using real-world data to improve immunotherapy response prediction 
• Closing care gaps through AI-enabled clinical workflows 
• The future of minimal residual disease, early detection and personalized cancer treatment 
• Why collaboration and data sharing will be essential to unlocking AI's full potential in oncology 

The views expressed in this podcast belong solely to the speakers and do not necessarily represent those of their organizations.

The views expressed in this podcast belong solely to the speakers and do not represent those of their organization.

If you want access to more future-focused, actionable insights to help biopharmaceutical companies better execute and succeed in a constantly evolving environment, visit the Syneos Health Insights Hub. The perspectives you’ll find there are driven by dynamic research and crafted by subject matter experts focused on real answers to help guide decision-making and investment. You can find it all at https://www.syneoshealth.com/insights-hub.

Like what you’re hearing? Be sure to rate and review us!

We want to hear from you! If there’s a topic you’d like us to cover on a future episode, contact us at podcast@syneoshealth.com. 

Transcript

SPEAKER_01 0:00

Welcome to the Synus Health Podcast, Early Signals, where we explore the ideas, innovations, and emerging trends shaping the future of oncology. I'm Dr. Y. L. Harb, head of RD and Scientific Strategy in Oncology at Synus Health. Recorded live from ASCO 2026, these conversations bring together perspective from leaders across biopharma, academia, and technology on the signals they're watching most closely, from AI-enabled drug development and real-world evidence to precision medicine. These conversations offer a snapshot of where oncology innovation is headed next. Let's tune into one of our conversations now. Today's conversation sits right at the intersection of two stories I've been watching closely. The immunotherapy revolution and the rise of AI in oncology. My guest helped write the first dose stories. As a physician scientist, Dr. Ezra Cohen was among those who defined modern checkpoint era in the head and neck cancer that reshaped how we think about who responds to immunotherapy and why. He's led precision immunotherapy at University of Chicago and UC San Diego. And he now serves as chief medical officer of Tempus AI, where he's working at the frontier of multimodal data and AI-enabled precision medicine. What makes Ezra a rare guest is the arc itself, the clinic to bench through the algorithms, now reshaping how we read biology. That's exactly the translation our early faced audience lives in every day. Ezra, welcome to Early Signals.

SPEAKER_00 1:41

Thanks, Waill. Thanks for having me.

SPEAKER_01 1:43

Take us from your through your journey, from Chicago to San Diego, from two major academic centers, now to one of the major sequencing companies and AI companies in the industry.

SPEAKER_00 1:56

Indulge me a little bit and let me start the journey a little bit earlier, and I think it'll give you perhaps an understanding for the basis for the career trajectory and the changes. I actually began to get interested in oncology as a family physician. I initially trained in medical school, and my first residency was to be a community family practice in a small town. And I did that. And as part of that, I was also doing home hospice, which of course 90% of our patients were oncology patients. And it was at that time when I began to get curious about the questions that we're still trying to answer today, but some we have answered. And those are the questions around why is this happening to this individual? Why have they or haven't they responded to treatment? Why did they get their cancer in the first place? Those types of fundamental biologic questions. And so I made a career change, going from a small-town family physician to re-entering training in first, obviously, internal medicine and then medical oncology. And that took me to Chicago and an interest, because the program is so strong at the University of Chicago, an interest in early drug development. And it also gave me the opportunity to work in the laboratory and start my own lab to bridge that divide between the basic science and the clinical and really open up a two-way street of finding discoveries in the laboratory that we could take to the clinic and then observations in the clinic that we could take to the laboratory. And I thought I would spend my entire career doing that at the University of Chicago. It was a wonderful place to work and exceptional colleagues. When I had an opportunity to take more of a leadership role at University of California, San Diego, looking at what was being built there, I really felt like I could make significant contributions. And so I moved to San Diego. And coincidentally, and a lot of the career success has been because I've been lucky at the right place at the right time. And coincidentally, immunotherapy was just entering oncology in a meaningful way with the anti-PD1 and other checkpoint inhibitors. And I found myself in San Diego that has an incredible breadth and depth of immunology, basic immunology, but not so much immunotherapy, which gave us at the cancer center an opportunity to build in oncology. And so that's exactly what we did. And again, things were going fantastically well. We had built a cell therapy lab, we had built a cancer immunotherapy program, a precision immunotherapy program. When the CEO of Tempest, Eric Lefkowski, reached out to me because we had been doing some work with Tempest, actually quite a bit, and said, we need a CMO. Uh-huh. And at first I thought I had no interest in leaving academia. But honestly, when I began to understand what Tempest was doing, it dawned on me that this is exactly what I have been trying to do my entire career, this type of translational research, blending molecular and clinical data and finding novel insights. But now Tempest was doing it at scale, at a scale that I could never even imagine. That prompted the third large move in my career, and that was to go to industry about three years ago.

SPEAKER_01 5:13

Very inspiring story, starting as a family physician and taking care of patients on hospice at the end of their life, dealing with difficult diseases like cancer, to motivate you to go back and study oncology, become medical oncologists, and doing lab research in immunology and getting into the clinical trials.

SPEAKER_00 5:34

The realization that really cancer is a molecular disease. And the reason I went into the lab is I felt that if we were going to make a difference, if we were really going to change the lives of cancer patients, we had to understand it at that level. If we didn't understand the molecular biology of oncology, then we're just throwing darts blindly at a dartboard and continuing to miss.

SPEAKER_01 5:55

We often think about precision oncology with biologically targeted therapy, small molecule. What's it strike me what you're saying? We really need to think about precision oncology in all therapeutics, including immunotherapy. Can you elaborate on that?

SPEAKER_00 6:09

Absolutely. I think any therapy, actually, this applies to any field of medicine, of course, but oncology is furthest ahead, mostly because, again, it is a molecular disease and we've had to understand it. When I think about precision, I think of it broadly. It's not just matching a molecule to a kinase that it can inhibit, but it goes way beyond that. It's really about okay, which patients with that mutation will benefit from that molecule? What type of side effects will that individual have? And how long could they potentially benefit? It really becomes the entire scope of removing uncertainty. And that is, I believe, the term that encapsulates everything that we do. We want to diminish the uncertainty in management and drug development as much as possible. And so it applies to immunotherapy. We have checkpoint inhibitors that, of course, are now approved in multiple tumors across the board. So there's a wide range of scenarios where these agents are effective. But if you look at the individual disease, I'll use head and neck cancer as an example because I know that very well. 20% of patients, even with our best biomarker selection, respond to an anti PD1 antibody. That means 80% of patients are not responding, and the great majority of that 80% are likely deriving no benefit. There's a lot of precision that can come into play there. We need better biomarkers. We need better understanding of the clinical characteristics of the patients. We need better understanding of which patients develop toxicities. And of course, we need better understanding of when we can start and stop therapy. All of these things go into that idea of precision oncology.

SPEAKER_01 8:01

At Tempest AI, you have been sequencing tumors for a long time and you have a very large data set and looking at multimodal analysis, whether looking at the DNA levels with genomics or RNA with transreptomics, and also looking at protein. This complex data set, and we're looking at the individual patient with the tumor profile. How are you able to predict response on that individual level?

SPEAKER_00 8:28

That's the goal. So I I have to admit that we are not there yet, at least not in the way we want to be, so there's still uncertainty. But to answer your question, let me take a step back again and say that that was the vision from the beginning, that the diagnostics, all of those things that you just mentioned, were a pathway to get data. Because the thought was that if we could combine molecular phenotyping through all those mechanisms that you talked about with clinical longitudinal annotated data, we can then begin to derive insight. So the diagnostics were really just a means to build the data. And then the data served as the foundation to create the AI. Tempest got lucky along the way because about five or six years ago, the large language models became significantly better. And we see that in our daily lives with things like ChatGPT and other agenc AI platforms. But those same large language models we can apply to a oncology database. So when you talk about how we can begin to understand the specific patient, well, right now it is based primarily on molecular profiling and whether that tumor has an actionable alteration. That started about a decade ago, and I think we're at a point where the low-hanging fruit has been picked. So I don't think we're going to discover many more drivers that alone can be targeted and yield a wonderful result. We have known drivers that we still need to understand how to target, and we saw a beautiful example of KRAS mutations in pancreatic cancer at the series ASCO. So we're getting better. I don't think we're going to discover very many more solitary drivers. Now, having said that, that's where the multimodal data come in. There are cancers that are thriving based on a multitude of genes, having knocked out specific tumor suppressors, having modulated transcription factors, and all of those things come into play in allowing that cancer to thrive. And so we begin to understand those things by combining the molecular phenotyping and the clinical, of course, using the AI, and we begin to derive some sort of metric to help us select which patients benefit from which therapies. A real-world example is our immune profile score. And getting back to the immunotherapy issue where you have checkpoint inhibitors, but we don't fully understand which patients benefit. So we were able to look at our database and find a discovery cohort of patients who had been treated with immunotherapy, either in first or second line, either monotherapy or combination, and then create a validation cohort that validated the initial findings, and we came up with the immune profile score, which actually outperforms true mutational burden, MSI status, and PDL1 expression. It's not perfect by any means, but we can now separate patients into IPS high and IPS low, and IPS high patients seem to benefit from immunotherapy and the converse for IPS low. So we're getting better. We're not where I want to be, but slowly but surely we're making those inroads.

SPEAKER_01 11:50

That's definitely a step forward. But how can we manage all this complex data where we're getting from each tumor where there's multiple mutations and trying to integrate that in prediction on the individual patient level?

SPEAKER_00 12:04

Absolutely, that's exactly what we're doing. And we just done that with EGFR inhibitors and non-small cell lung cancer. We had a press release just a few days ago. This was the first output from our oncology foundational model. Basically, it does exactly what you said for a patient with non-small cell lung cancer that has an EGFR activating mutation. It can take that patient, look at all the co-mutations, look at other variables, including some clinical characteristics, and determine whether that patient is likely to benefit from an EGFR tarzin kinase inhibitor or not. Not to the point of 100% versus zero, but with a much greater degree of certainty than what it would be, which is about 50-50.

SPEAKER_01 12:47

That's exciting. Let us know about what other exciting things you're working on and you think is going to change how we manage a patient with advanced cancer or early cancer for that sick.

SPEAKER_00 12:56

I think it's both. I think we'll hopefully get into AI applications to early detection and prevention, but we as a company are not there yet. In terms of earlier stage cancers and advanced cancers, there's a lot that we're trying to put in place. And I'll just talk about a couple of examples because otherwise you'll be listening to me for a couple of hours. One example in early stage cancers, we're looking, as many companies are at minimal residual disease testing, the idea that we can, with ultrasensitive tumor-informed assays, begin to change management. We're in the early days of that. We still need to collect more data, but again, we can use multimodal data to create those predictions. But even more exciting is what we're doing for advanced patients now on a routine basis using AI. For instance, we've deployed digital pathology AI to, on the front end, give the provider an early look with respect to what alterations may be present in that tumor just from the HE. So that answer comes back within 48 hours, and we can say with a high degree of probability, in the 90% range, whether that tumor has, for instance, an EGFR mutation or a FGF receptor mutation and things like that. And then on the back end, we're using digital pathology AI again when a tumor doesn't have enough tissue. So for whatever reason, we were not able to do the sequencing. We can use Ditch Path to say, okay, we didn't sequence this patient, but the HE is telling us that there is this alteration with about an 83% certainty. I'm making that number up, but a high degree of certainty. And then the provider can decide whether they want to get another tissue, they want to do an alternative test, et cetera, et cetera. So that's the AI as it applies to digital pathology. Also, something that's very exciting, and we presented these data at this meeting, we can bridge care gaps. We all know the NCCN guidelines, and we all would love to think that we follow the NCCN guidelines for every patient that we see. But the reality is that we're very busy providers, and the guidelines are for non-small cell lung cancer over a thousand pages, and it may be difficult to adhere to those. So, what we created was bridging care gaps by embedding deeply into the EMR, understanding which patients are not following guideline direct care, and then surfacing that to the provider. In the first instance, we did this with EGFR mutational testing in early stage non-small cell lung cancer. We've done that with ALK testing, we've done that with PDL1. We have similar care gaps in breast cancer and prostate cancer. The idea is that we help the provider at least follow, and if they want to, follow guideline direct care. And in fact, what we've seen is what we call a lift of anywhere from 15 to sometimes 30 percent in that management related to what the guidelines are saying. So it really works. And this makes a tremendous difference, of course, to patients' lives.

SPEAKER_01 15:56

That's exciting with pathology, imaging AI able to predict some molecular levels changes that we might have missed.

SPEAKER_00 16:03

Yeah.

SPEAKER_01 16:04

And trying to close the gap in implementing approved treatment for patients. You mentioned gaps, and so I'm gonna use that theme. And where do you see the gaps in ability to predict what tumor responds to what treatment or combination of treatment? Is it the size of the data? Is it the correlation with clinical information? Is there understanding the biology? You mentioned you're not quite there yet. What the roadmap looked like to get there?

SPEAKER_00 16:32

I think it's all of those, and I would add compute power to that. The size of the database is critically important. We have a very large database, but it also, as new therapies come up, because we depend on real-world data, in certain cohorts we have to wait for that data to accumulate. So the size of the database is critically important. The quality of the data, the diversity of the data, the generalizability of the data is incredibly important as well. Fortunately, with AI, we now are able to take unstructured data, so progress notes, all types of reports in different formats, and we can synthesize it, we can interrogate it without having to do that manually. So the AI has helped tremendously there. But that's always getting better and getting faster. So there's a bit of a gap in terms of what the AI can do and what it might do even in the next six months, let alone the next two years. Then there is the question of how do we then understand the biology, a better fundamental understanding of the biology of the tumor and the host, as well as how they interact with the different therapies, is also critical. And that means we have to go beyond real-world databases and go into other types of models, either preclinical or even clinical models. The last part that perhaps is the biggest challenge right now is implementation. We can build all these wonderful tools, but if we can't get them to patients, if we can't embed them in the workflow of a provider, then they'll never be used. Even if it's wonderful, it will never be used. So part of the challenge is making sure that providers, and for that matter, patients, have easy access to this information exactly at the time that they need it. What we've tried to do is embed into the electronic medical systems in a way that the provider gets notified around the time that they're seeing the patient, so they have that information at hand. And we're building tools that are patient-facing to help them understand their disease at a much deeper level and ask their provider the appropriate questions that are specific for them and the tools to understand their cancer at a much deeper level.

SPEAKER_01 18:46

Indeed, understanding the cancer at much deeper level would enable us to have better treatment.

SPEAKER_00 18:51

I hope so.

SPEAKER_01 18:52

Talking about early phase, especially when we, as an oncologist, we know that most cures happen in early phase cancer, so early detection is a key, but also knowing what is the best treatment and how to tailor it for that tumor and monitoring response. I would like you to comment on where the field is going in using molecular data, in making the treatment more tailored to that tumor and patient needs.

SPEAKER_00 19:17

We have a long way to go. My dream is hopefully I'll realize it in my practice lifetime, is we do exactly what you've just described. We take a patient's tumor, we look at the patient themselves as a whole, and we begin to tailor their therapy with the best tools that we have to make it the least toxic with the highest chance of survival. We're beginning to see this in small pockets, but you can imagine a day where there may be patients who could be cured of their cancer with just immunotherapy. We look at MSI high rectal or colorectal cancer, and we see that some of those patients in early stage are actually cured, as far as we can tell, with just immunotherapy alone. If we had even better biomarkers across tumors, maybe we can extend those findings to many other patients. There may be patients who benefit if they're getting radiation from much lower doses of radiation. Remember, most radiation protocols were developed 50 years ago mostly by trial and error and not based on a solid fundamental understanding of biology. And I don't want to diminish it. There was a lot of understanding and a lot of research that went into it, but I'm confident that we can cure many patients with much lower doses. And then one wonders whether patients need the specific types of surgeries that they're getting, and maybe we could modify those, especially ones that affect function, long-term function, to reduce dramatically long-term toxicity. But the challenge is that especially in the curative setting, all of this takes time. Because the last thing as a field we would want to do is diminish survival. As we begin to reduce therapy or specify therapy, and sometimes it's increased the intensity as well. In the curative setting, we have to do that through prospective clinical trials. We have to prove that we are still delivering the best outcomes in a systematic way because the last thing we'd want to do is diminish survival.

SPEAKER_01 21:18

Can you comment on how we can collaborate on a wider scale? Because you mentioned that there's limitation in the data size, data quality, which means that we really this needs to be collective effort. Are we structured to be able to do that? Is there an obstacle for us as a community, as an industry, as medical community to be able to collaborate and aggregate this data to be meaningful in making these decisions?

SPEAKER_00 21:47

Honestly, I think it's inevitable. The short answer to your question is yes. The barriers are really around collaboration and opening things up in a way that still protects patients' privacy and all those things that we Really hold dear, but that's where the AI is really helping us. There was a time in my career where I thought we would never get there unless all physicians structured their notes the same way. But that's no longer the case. Now with large language models, we can take completely unstructured data and make sense of it. Even when physicians use short forms or language that can be idiosyncratic, the AI can decipher it. So that problem is being solved. The bigger issue is getting around this idea of sharing data. I think it will happen. I think it's inevitable. I think it's in the best interest of patients and getting novel therapies to cure diseases much faster. But we have to break those cultural barriers more than the technical barriers.

SPEAKER_01 22:45

That sounds really promising. What excites you at this ASCO? What have you seen or heard that really gets you excited?

SPEAKER_00 22:53

I have to say the plenary session at this year's ASCO is one of the best in terms of changing practice that I remember going to. Every single one of those abstracts, we're talking on Monday, and as of 8 a.m. today, in all of those diseases, my practice changed. Granted, all the drugs aren't approved yet, but I'm sure they will be based on the data that we saw. But in pancreatic cancer, in sarcoma, in non-small cell lung cancer, right away, these are practice-changing standard of care types of findings. And it's incredibly exciting.

SPEAKER_01 23:28

It's very exciting when we see some practice-changing data that would help us give better treatment for our patient, indeed, especially in pancreatic cancer, where there has not been any good treatment for decades.

SPEAKER_00 23:38

I mean, amazing. There was a standing ovation. Yes. This is incredible.

SPEAKER_01 23:41

Five years from now, where do you see where we're going?

SPEAKER_00 23:44

I think really in five years, and this speaks to how quickly things are happening. I really do believe that in five years, a clinician will be able to see a patient and they will be able to tell them with a much higher degree of certainty. So right now, many of our therapies have 20% response rates, 30% response rates. Optimistically, with more targeted therapy, 50% response rates. But I think we'll get to a day where we'll be able to tell a patient, you have a 90% chance of responding to this. These are the most common side effects that you might have based on your pharmacogenomics. I'm going to give you this wearable so that if you begin to have side effects, and I'm going to give you this tool on your phone that you can alert us before the side effects get severe, so that we can intervene. And if and when that drug stops working, I know what I'm going to do as the next step with a high degree of certainty. I do believe that within five or ten years in oncology, four things will happen to patients. We will detect a cancer, we will prevent a cancer, we will detect a cancer early where a simple intervention is curative. We will cure many patients with advanced cancer, and we will render the other group of patients with essentially a chronic disease. And I think we'll see that in our lifetime.

SPEAKER_01 25:03

These are all great predictions and very helpful for cancer patients. Ezra, any final thoughts?

SPEAKER_00 25:10

I've been doing this for uh over 20 years, and I have never been as excited as I am right now about being in oncology. It's just an amazing time.

SPEAKER_01 25:18

Well, Ezra was exciting to have you on our Early Signals podcast. Very wonderful conversation and very insightful and hopeful for our patient. Thank you very much for joining us. Thank you. It was a pleasure. That's all for today's episode of the Sinious Health Podcast, Early Signals. I'm your host, Dr. YL Harb. If you have other topics you would like to hear on this podcast, please send us a message at podcast at seniorshealth.com. For access to more future focused, actionable life science insights, please visit the insights hub at sineushealth.com.