Predicting Immune-Related Adverse Events Through Machine Learning With OncoHost: Dr. Waqas Haque, Dr. Matthew Hadfield, and Dr. Ofer Sharon

In this first episode of Oncology Data Advisor’s new podcast series, “Exploring Artificial Intelligence in Cancer Care,” Dr. Waqas Haque and Dr. Matthew Hadfield of the OncData Fellows Forum engage in a discussion with Dr. Ofer Sharon, CEO of OncoHost. Dr. Sharon provides an overview of the clinical need for increased knowledge of patients’ probabilities of developing immune-related adverse events and explains the PROphet® platform, OncoHost’s novel plasma-based, proteomic pattern analysis tool that uses a single blood sample to guide immunotherapy treatment decisions.

Keira Smith: Thank you, everyone, for tuning in to the show today. We’re so excited to be hosting this first episode of Oncology Data Advisor’s new podcast series focused on artificial intelligence (AI) in cancer care. I’m joined today by our two co-hosts from the Oncology Data Advisor Fellows Forum, Dr. Waqas Haque and Dr. Matthew Hadfield. We have the honor of being joined today by Dr. Ofer Sharon, who is the CEO of OncoHost. We’re really looking forward to hearing more about OncoHost and your work in this field. To start off, we can have everyone go around and introduce themselves. Dr. Haque, I will turn it over to you to start off the introductions, and then you can take it away.

Waqas Haque, MD, MPH: Awesome. Thanks so much, Keira. Good morning, everybody. I’m Waqas Haque. I’m a third-year Internal Medicine Resident at New York University in a Clinical Investigator Track, and I recently matched to the University of Chicago for Oncology Fellowship. I’m interested in melanoma research and immunotherapy, and also in looking at new applications for AI in predicting adverse events. I’m really excited about this podcast today.

Matthew Hadfield, DO: I’m Dr. Matthew Hatfield. I’m a third-year Oncology Fellow at Brown University in Providence, Rhode Island. I’m in the process of transitioning into being an Attending Junior Faculty Member at Brown, with a focus on early drug development and phase 1 clinical trials, as well as cutaneous malignancies including melanoma. My research and clinical interests are predominantly in immunotherapy, both the development of new immunotherapeutic medications for patients as well as managing and diagnosing immunotherapy toxicities, with a particular interest in developing biomarkers to predict immunotherapy toxicities. It’s very nice to have you, Dr. Sharon, and I’m really looking forward to this conversation. I think this is such an important topic that we all can benefit from learning more about.

Ofer Sharon, MD: Great. Thank you guys, and thanks for having me. It’s a real pleasure, and I’m always happy to speak about the work that we do. I’ll give a short introduction about myself. I’m Ofer Sharon, the CEO of OncoHost. I’m an Internist by training, and in the last 25 years, I was basically founding companies. OncoHost is my third company. My passion is the combination of mathematics and medicine, to put it very simply. We used to call it big data; now we call it machine learning, but it’s all a similar approach to finding clinical applications in very complex situations where you have a lot of data coming from a lot of sources and you need to make sense of it. One company that I was involved in was sold to Baxter in 2018. I have another company that is currently operating out of New York, which I co-founded almost 10 years ago. OncoHost is the latest company.

The idea for OncoHost came from the time I spent with Merck. I was the Head of Medical Affairs for Merck in our region. I was very lucky to be part of the Global Development Team of pembrolizumab (pembro) which is the second blockbuster immunotherapy that came to the market in the first two indications, melanoma and lung. We saw an amazing promise in the drug. We thought it would almost be a universal drug for cancer because it’s not targeting the cancer; it’s targeting the immune system. The adverse event profile was very different from what we’re used to from chemotherapy or targeted therapy. There was a lot of hope with it. Ten years later now, we have come to the realization that this treatment is limited, and we do need to find a way to optimize how we manage our patients using those new modalities. This is why I decided to start OncoHost.

Dr. Hadfield: That’s a wonderful introduction, Dr. Sharon. I think it really segues well into this conversation. To your point, when the first immune checkpoint inhibitor, ipilimumab (ipi), was studied in melanoma, we went from an overall survival of eight months to three years with just ipilimumab. Then the combination of ipi and nivolumab (nivo) together pushed 60-plus months in median overall survival. But on the opposite side of that, with all the hope that came with immunotherapy, we certainly have more questions now than we did 10 years ago and a lot of things that still need to be explored.

One of the biggest things is that when we look at things like combination immunotherapy, particularly ipi/nivo in metastatic melanoma, half of those patients will develop an immunotherapy-related adverse event. I think you have such an interesting perspective of being in the development of pembro, which is, to your point, a blockbuster drug and really the backbone of our first-line treatment for non–driver-mutated non–small cell lung cancer. Maybe you could walk us through a little bit about, for the broader audience, immunotherapy—how it works and how immunotherapy-related toxicities differ from target therapy or chemotherapy, which is something we’re a little bit more familiar with.

Dr. Sharon: Right, of course. So, it all starts with the mechanism of action. The modern immunotherapy drugs are basically intervening in the way the immune system is reacting with tumor. Potentially, the immune system holds the ability to kill cancer. To put it very simply, those cells can identify tumor cells as foreign and attack them. But we all know that cancer is a very sophisticated entity, as I like to call it, in a way. The ability of the cancer to hide from the immune system is a very interesting phenomenon. We see that cancer cells are basically able to mimic normal cells, so the immune system is not attacking them.

The beauty of the immunotherapy is the ability to unveil this camouflage that the cancer cells are using and making the cancer cells vulnerable to the immune system. Here we have a treatment modality which is not limited to cancer type. We can use it in many cancer types. Matthew, if you remember the early days, we started looking at what are now called basket trials, where you have very small cohorts of patients in a lot of indications. The reason and the logic behind it are very simple. Why limit ourselves to one cancer type if we can test in multiple types at a time?

What is interesting is that in parallel to the efficacy that we saw in some patients—and we need to get back to this concept of some patients later when we talk about adverse events as well—we also started to see adverse events that we were not unfamiliar with. Let’s consider for a second what we’re familiar with from chemotherapy, for example. Chemotherapy, if I take a military metaphor, is like artillery. Potentially, it kills every cell that multiplies at a high pace, no matter whether it’s a normal cell or a cancer cell. With that, of course, come the adverse events that we’re familiar with—hair loss, disease of the gastroenteric system, the bone marrow. Whenever you can find a sensitive, high replication rate, you will see adverse events.

Some of these may be devastating, and some may be very uncomfortable for the patients. But this is what we saw, and I think this is one of the biggest fears when you speak to patients about cancer and chemotherapy. This is what they’re afraid of—”I’m afraid of losing my hair. I’m afraid of all those adverse events that I’ve heard about.” It’s very visible to the eye. Immunotherapy is different. Immunotherapy is not always visible to the eye because the adverse events of immunotherapy are exactly what the drug is supposed to do—activate, prime the immune system against cells. In some cases, the cells are losing their way and attacking normal cells.

That poses an interesting challenge because at first, it was very hard to differentiate between immune-related adverse events and just adult diseases. I’ll give you a very interesting example. A patient that I treated came to the emergency room with symptoms of myocardial infarction (MI), chest pain, and radiation to other areas. He was cold and sweaty. When we asked him if he had any background illness, the answer was, “No, I’m healthy.” He forgot to mention that he’s a cancer patient. He had lung cancer, in this case, treated with immunotherapy. Cancer is a different entity for some of those patients. It’s not part of the chronic disease. He was diagnosed and treated as a patient with myocardial infarction, and he unfortunately passed away during catheterization.

Postmortem, we learned that the patient was not suffering from myocardial infarction. He was suffering from an immune-related adverse event which mimicked the myocardial infarction. It’s called myocarditis or pericarditis. The treatment, of course, is very, very different. It’s relatively simple, but definitely treatable if you can think about a diagnosis. This is very tricky and very challenging for some clinicians to make sure that when you see something, when you see an illness in a patient that’s treated with immunotherapy, you should consider the immune-related adverse events. The immune cells are attacking normal cells. You can see disease of the thyroid gland; you can see diabetes; you can see, as we mentioned, myocarditis and pericarditis; you can see neural diseases. Basically, every disease that we know of that we call an autoimmune disease is something that we can see as an immune-related adverse event.

Dr. Haque: Thanks so much for sharing that, Dr. Sharon. I really appreciate the example you mentioned about the patient with the MI who didn’t mention his cancer history. Just for the broader audience, can you talk a little bit more about the time course for which we see immune adverse effects and some of the ways that oncologists manage them?

Dr. Sharon: Interestingly, with immunotherapy response, it’s not always evident immediately. Sometimes it takes some time until we see response. With immune-related adverse events, the picture is similar. In most of the publications I read of patients treated with immunotherapy, usually you see the major adverse events in the first three months of therapy. It doesn’t mean that adverse events cannot appear later. They can and they will, but in most cases, you see those in the first three months of therapy. I think this is where you need to be, as a clinician, on your highest alert, making sure that what you see should be considered as a potential immune-related adverse event.

With any adverse events, we are treating cancer. As clinicians in medicine, it’s all about risk-benefit. We’re willing to risk even severe adverse events in order to offer the patient potential cure in earlier stages of disease or elongation of life in advanced diseases. I think that when we think about immune-related adverse events, focusing on, let’s say the first three to four months of therapy, we should think about those that may cause treatment change or discontinuation. We should be alert to those. The patient should be alert to those.

When we are talking to our patients, we should tell them, “Listen, this is what you might experience in the next three months. Don’t ignore it. If you have a headache, your vision is blurry, or your blood tests are not as usual, don’t ignore that. Come to me. Tell me about it. Let’s make sure that what we are seeing is not an adverse event that we can potentially treat.” At least in my mind, I think this is the most important part. For us at OncoHost, this is also the main focus. We’re focusing on what we call the significant adverse events, grade 3 and above, those that might cause treatment end or change, things that are meaningful to the clinician on the other side of the reports.

Dr. Hadfield: We’ve hit on so many really important points here. I think a few things that I would highlight from all the information that you’ve gone over is at the current time, we really don’t have any great predictors for who’s going to develop a toxicity. If you’ve practiced oncology, if you’ve been involved in the care of cancer patients, you’ve lived through this horrible situation where you start someone on an immunotherapy and you talk to them. There’s this huge misconception that immunotherapy is chemo-sparing, which I think really is a very big problem when you speak to patients because a lot of patients interpret that—and there are a lot of news stories in the lay media about immunotherapy that essentially make it seem as though they’re without toxicities, and that could not be further from the truth.

At the current time, we don’t have great predictive biomarkers. There have been some really great studies out of Dana-Farber and Massachusetts General Hospital (MGH) that looked at human leukocyte antigen (HLA) subtyping, but so far, those haven’t really panned out to predict exactly who’s going to develop, as you mentioned, these severe toxicities. For me, the thing that really scares me and gives me more urgency, as we’re going to talk about more ways to predict toxicities, happens as these therapies start to trickle down to the neoadjuvant setting. This is something that I’ve talked about with a lot of people before. In the metastatic setting, you can shift that balance more towards, “I’ll accept more toxicity because we don’t have a great option for your cancer.” Unfortunately, it’s a terminal illness. But when you start thinking about giving pembrolizumab in the neoadjuvant setting in melanoma, as we’re doing now based on the recent studies done in that disease space, you’re going to give someone a toxicity that’s eventually going to lead to them not being able to get surgery or be fatal. You could have taken away a curative-intent therapy with a toxicity. The urgency here is very pressing.

I would love to talk briefly, just to round out the conversation on immunotherapy-related adverse events—what, in your experience, has been the best way to manage these toxicities? I know mostly we rely on steroids, and once we get past steroids, it’s case series–level literature to manage these toxicities. Then maybe we can start hearing a little bit about OncoHost and how that’s going to hopefully add to our knowledge base of how to predict these types of toxicities.

Dr. Sharon: I think that like many, many things in medicine, this is also unfortunately something that is very interesting in terms of mechanism of action and why the patient is suffering from specific adverse events. But when it comes to actually treating the adverse events, we don’t have great solutions. You mentioned the management, which is steroids. In some cases, you need to stop therapy completely or delay it. These are the best options. Like you said, when we go to the extremes, you will see some experimental drugs that are more daring in cases where we are facing very severe adverse events up until the point where there’s a risk for the life of the patient. I think one of the interesting cases, and given your specialty, that we read about is Stevens-Johnson syndrome. It’s not common, but it’s deadly if not identified quickly and treated the right way.

I also agree with you about the concept of the level of concern, because at the end of the day, this is how we manage patients. Like I said earlier, it’s risk-benefit. The drug I’m giving you should give you more benefit than harm. This is what we want to do. This is how we manage our patients. In a way, we are playing a game of probabilities. The probability of getting a benefit after the drug should be much higher than the risk of an adverse event. I’m not going to go into treatments because like I said, it’s mostly anti-inflammatory drugs. We are talking about steroids, and I’m not going to talk about the science fiction in some of those cases, the most severe cases.

But I really want to relate to your point about the adjuvant and neoadjuvant setting because this is where we shift the focus in a way from efficacy to toxicity. There should be much more emphasis on toxicity because we have choices. In the metastatic setting, sometimes we don’t have a choice and we need to do whatever we can. We are willing to pay the price. I think this is where it gets really interesting. The way we decided to look at it, because OncoHost is a company that is dealing with probabilities, is probability of clinical benefit, and with the same concept, with the same algorithms, with the same approach, but with very different mechanisms of action, the probability of adverse events.

When we are developing a product, we try to start with a clinical question. The clinical question of efficacy is relatively straightforward. We have three clinical outcomes: the response rate, the progression-free survival (PFS), and the overall survival (OS). Generally speaking, we are trying to create an algorithm that will provide the clinician with a prediction for a meaningful clinical outcome. We think about how we are going to train the algorithm and what it is going to predict. If I’m going to give you, as a clinician, a prediction for a minor headache, there’s not a lot I’m doing for you here. What we needed to do is start by identifying the clinical need, and the clinical need, the way we understood it by talking to clinicians, is to have a predictor for the severe adverse events or the significant adverse events.

How do you define those? There are very different definitions for severe adverse events in clinical trials and for severe adverse events in real life. Clinicians told us, “We want organ-specific predictions.” We can name the organs—the brain, the heart, the thyroid, and the pancreas are, I would say, the most significant organs that might be affected. We have the organ-specific adverse events, but then we also have the grade 3 and above adverse events. Those are the ones that cause the patient a lot of suffering, sometimes to a point where the patient will tell you as a clinician, “Listen, I prefer to stop therapy.” You say, “Hey, but we are treating your cancer.” “But I cannot tolerate the drug.” Severe adverse events of grade 3 and above, organ-specific adverse events, and then adverse events that may cause treatment end or change—this is how we started by defining what we are training our algorithm for.

Dr. Haque: Got it, thanks much for sharing, Dr. Sharon. In terms of the solution that you helped create, can you talk a little bit about the recent clinical trials, some of the results from that, and maybe if you can also touch on how providers can explain this to patients?

Dr. Sharon: Sure, so let’s start with the clinical approach and then go towards what we do at OncoHost. What we’re trying to do at OncoHost stems from the understanding that biomarkers for immunotherapy are different from the standard biomarkers that we know of. I think the best example is genomic biomarkers. Genomic biomarkers are relatively simple. It’s a matter of “on” or “off.” The cancer cell has a genetic mutation that we can target or not. It’s almost a generic question: yes or no.

The immune system is a little bit more sophisticated. We are talking about different cell types. You’re talking about the interaction between the cells and the cancer, between the cells and the host body. We are starting to think about the concept of multicomponent biomarkers. In order to identify multicomponent biomarkers, the most logical place to start our search was the blood because the blood, in a way, is a soup where you can see all of the biological processes that are happening in the body. That comes, of course, with a lot of noise because you have a lot of processes. Some of them are not necessarily related to the tumor, and some of them are not related to the direction between the tumor and the therapy.

I would say that the first focus was to identify those proteins that we believe are relevant to the treatment. Out of those, once you identify those proteins, you need to start thinking about the signaling pathways, the biological processes that are taking place in our part of the response or lack of response of the patient to treatment, and also the adverse events or lack of adverse events. This is a very interesting game. It’s almost like playing a game where you are trying to be a detective and find those biological processes that are involved in the adverse events. We’re identifying proteomic patterns, we’re looking at patient groups, and we’re looking at the difference in protein expression between patients that are responding to treatment and those that are not responding to treatment, and between patients who are suffering from adverse events and patients who are not suffering from adverse events.

We’ve identified proteins with very significant difference in expression level between those patient groups. We term those “toxicity-associated proteins” when it comes to identifying adverse events. The next step is trying to identify how many of those proteins are overexpressed in each of the patients. We saw that among patients who have a lot of those proteins overexpressed, and I’m keeping it simple on purpose, the patient is likely to suffer from immune-related adverse events. With few of those toxicity-associated proteins, the patient is less likely to suffer from immune-related adverse events. We saw with an accuracy of about 92% that we can actually predict if the patient will suffer from a significant adverse event, as I mentioned earlier, grade 3 and above, and those that may cause treatment end or change even before we start therapy by identifying those differences in protein expression in the plasma.

The result has two parts to it because this is an algorithm-based product. The machine or the algorithm is developed on a certain cohort of patients, and we call this one a training set. Then we test the ability of the algorithm to repeat the result on a new cohort of patients. The first thing you discover is the power of the algorithm, the ability to predict for adverse events. Then the next step is trying to make sense of it and understand which mechanisms of action are relevant, which adverse events we can expect to happen.

I have to tell you that I don’t see this as a standalone tool. I think that when I want to create a product for a clinician, I need to give the clinician the full picture, namely the efficacy or the probability of efficacy, but also the probability of the adverse events, because this is exactly what Matthew said earlier. As a clinician, especially in the early stages in the adjuvant and neoadjuvant setting, I want to be able to take decisions based on risk assessment: how much benefit versus how much risk for my patient. Even if I do not consider the immune-related adverse event a huge risk, at least I can be ready for it and tell my patient, “Listen, according to this test that I see, there’s a very high likelihood that you’ll suffer from an immune-related adverse event. This is what you should expect. Don’t hesitate to give me or my nurse or my office or my clinic or a call. If you feel that you’re not okay, we can identify it earlier, we can treat you earlier, and we can fix it.”

I think this is also the way we should explain it to patients. “The fact that you see in the report, ‘highly likely to have adverse event,’ doesn’t necessarily mean that you will suffer from an adverse event. All it’s saying to us as clinicians is that there is a probability that you will suffer from an event. This is actually a great benefit for me as a clinician because I know I should be ready, I know I should prepare, and I can decide on different regimens of treatment based on this risk assessment.” I try to really simplify it as an additional lens. We’re providing the clinicians with a tool to improve the resolution of the story. I’m seeing the patient, the physical examination, the blood test, and the imaging. Here I have another lens telling me something not only about the ability to benefit from treatment, but also about the risk of developing significant adverse events.

Dr. Haque: Just so I can understand correctly, it seems like the algorithm is mainly based on protein-based biomarkers, and that’s where you’re getting the prediction from. We definitely want to congratulate you and OncoHost on getting the 92% accuracy rate. That obviously sounds very promising for patients in the long term. Just sort of thinking about wanting to do better and thinking about that 8% more that we have to go, have you considered other patient covariates—body mass index (BMI), age, functional status? Are there radiological findings or other things we can incorporate into the blood-based biomarkers to maybe increase the accuracy?

Dr. Sharon: That’s a great question and also a very significant caveat that we, as a producer of a product, especially a machine learning–based product, should be concerned about. Do I see an actual signal from the protein, or do I only see the differences in the BMI, sex, or age of the patient? I can promise you that we ran multivariate analysis, and we ran univariate analysis. To date, we do not see any correlation with the clinical feature that can explain the probability of the effect or the ability to predict an adverse event. I would say, generally, the older the patient is, the higher the risk is. We do also see some correlation with BMI.

Other than that, I think that you will see papers talking about differences between men and women. We see those as not significant, not a clear signal that we can identify. But relating to what you said, I think that’s a very interesting next step. It’s definitely going to be part of the research and development (R&D) process for us—trying to improve the result and the accuracy, and get even higher numbers by combining the algorithm with a potential clinical feature. This is still a work in progress. It’s an interesting question, and it remains to be seen whether we can improve on that.

Dr. Hadfield: That’s great, Dr. Sharon. I just want to reiterate again, thank you so much for joining us today. I think I couldn’t overstate enough that developing methodology for predicting immunotherapy-related adverse events—particularly, as you pointed out, the more severe adverse events—is not only going to be helpful, but it’s really going to be a critical part of how we manage these patients moving forward. Particularly as we develop new immunotherapies and new immunotherapy combinations, we can actually sit down with a patient, look them in the eye, and tell them, “Hey, I think based on this, you have a higher probability of getting a toxicity.” We’ll be able to help them make a more informed decision on what they want to do and what amount of risk they want to accept. I think this is just fascinating material, and this has been a wonderful conversation.

Dr. Sharon: Thank you. It was a pleasure speaking with you guys, and thank you for the challenging questions. I hope I did a good job answering those.

Keira Smith: This was fantastic. This was an awesome conversation, and it’s so great to hear about this work and its potential for predicting immunotherapy-related adverse events. Thank you so much again, Dr. Sharon, for coming on the podcast today, and thank you to our two co-hosts for leading a great conversation.

About the Speakers

Waqas Haque, MD, MPH, is a third-year Internal Medicine Resident at New York University in a Clinical Investigator Track. As a Clinical Investigator Track Resident, Dr. Haque has balanced his patient care work with a variety of research projects. He hopes to begin fellowship training next year in Medical Hematology/Oncology at an academic program with opportunities to further his work in innovative clinical trial design, value-based care delivery to cancer patients, and becoming an early-stage clinical investigator.

Matthew Hadfield, DO, is a Hematology/Oncology Fellow at Brown University/Legoretta Cancer Center in Providence, Rhode Island. His research focuses on melanoma and early-phase clinical trials, and his main areas of interest include early-phase drug development, novel immunotherapeutic combinations to overcome therapeutic resistance, and predictive biomarkers for immunotherapy toxicities.

Ofer Sharon, MD, is a physician and entrepreneur with over two decades of experience in clinical research, pharmaceuticals, and biotechnology. He has made vital contributions to the acceleration of personalized medicine and oncology drug development. Dr. Sharon is the CEO of OncoHost, a technology company transforming the approach to precision medicine for improved patient outcomes. Prior to joining OncoHost, Dr. Sharon served multiple roles in global pharmaceuticals, including AstraZeneca and Merck, and he was part of the pembrolizumab clinical development and launch team. He has co-founded several health care companies centered on bioinformatic and machine learning platforms for clinical deterioration detection and early intervention.

For More Information

OncoHost (2023). PROphet®. Available at:

Naidoo J, Reinmuth N, Puzanov I, et al (2023). Pre-treatment plasma proteomics-based predictive biomarkers for immune related adverse events in non-small cell lung cancer. J Immunother Cancer (Society of Immunotherapy for the Treatment of Cancer 38th Annual Meeting Abstracts, 11(suppl_1). DOI:10.1136/jitc-2023-SITC2023.1229 

Transcript edited for clarity. Any views expressed above are the speakers’ own and do not necessarily reflect those of Oncology Data Advisor.

Related Articles