Using Dynamic Models to Inform Treatment and Prognosis in AML

Podcast Transcript

Dale Shepard, MD, PhD: Cancer Advances, a Cleveland Clinic podcast for medical professionals, exploring the latest innovative research and clinical advances in the field of oncology. Thank you for joining us for another episode of Cancer Advances. I'm your host, Dr. Dale Shepard, a Medical Oncologist and Co-Director of the Sarcoma Program at Cleveland Clinic. Today, I'm happy to be joined by Dr. Moaath Mustafa Ali, a Hematologist and Medical Oncologist here at Cleveland Clinic. He's previously a guest on this podcast to discuss the impact of p53 in patients with acute lymphoblastic leukemia. That episode is still available for you to listen to. Today, he is here to discuss models to predict progression of patients with AML, so welcome.

Moaath Mustafa Ali, MD, MPH: Thank you. Thanks for inviting me again.

Dale Shepard, MD, PhD: Yeah, absolutely. So, remind us what you do here at Cleveland Clinic.

Moaath Mustafa Ali, MD, MPH: I'm a hematologist, a medical oncologist, specialized in treating myeloid neoplasms, an acute leukemia specialist. I treat patients with AML, acute lymphoblastic leukemia, other blood cancers, but I also do a lot of observational studies, including in myeloid neoplasms and even in solid tumors treated with immunotherapy, and I work also part of a lab, Tim Chan lab as well.

Dale Shepard, MD, PhD: Excellent. So, pretty wide ranging and as part of that wide experience, we're going to talk about some modeling things. We're going to talk about how sort of use clinical and genetic and treatment sort of data, and come together with some modeling. Sort of, I guess as a backdrop, a lot of different people might be listening in. We're going to talk about AML and why has that been particularly a difficult disease to sort of predict responses or prognosis?

Moaath Mustafa Ali, MD, MPH: First of all, acute myeloid leukemia is one of the most aggressive cancers that hematologists and oncologists deal with. It is characterized by acute presentation and, unfortunately, short survival, and it's a disease that is very dynamic because it's based on a blood cancer, so you would imagine it's very dynamic, and hence it can change the driver mutations quickly, which can result in relapses quickly, as well as sometimes even refractory diseases from the beginning, and hence it becomes a more difficult disease to predict. Also, you don't have much time. You have to make decisions quickly for this disease.

Dale Shepard, MD, PhD: And I guess, just again, for a variety of people listening, this is really, when you say make decisions quickly, we're talking really quickly. Give us a perspective of when, from diagnosis to treatment, what's the projected start time?

Moaath Mustafa Ali, MD, MPH: Most of the time of acute myeloid leukemia, you generally have to make a decision within one to two weeks. Sometimes you actually have to make a decision overnight. A lot of times I have patients get admitted to the hospital. We have to make a decision overnight, and we have to consent patient for chemotherapy and start chemotherapy overnight. Frequently, it can be very urgent, and these decisions sometimes can be rushed, and a lot of times it can be difficult decisions.

Dale Shepard, MD, PhD: And when you talk about this disease and looking at mutations and how you select therapies, sometimes you just don't have the time to start that.

Moaath Mustafa Ali, MD, MPH: A lot of times, we have to start treatment even without knowing the full picture of the AML. Sometimes we have to start, for example, ATRA, all-trans-retinoic acid, even in patients we don't have a confirmation of the diagnosis, since these complications can be lethal, and hence rapid intervention is necessary.

Dale Shepard, MD, PhD: And so, we're going to talk about this modeling, kind of how you got there, and what we found so far with that, but what do we do currently in terms of what do we use right now to predict a response or to therapy and how to predict prognosis?

Moaath Mustafa Ali, MD, MPH: The science of acute myeloid leukemia has really transformed over the last 10, 15 years. In the past, we used to use histopathology to classify AML subtypes. More recently, we started using cytogenetics and mutational data, probably even more than other cancers. This wealth of knowledge has helped us in prognostication. However, the prognostication systems that we use currently, European Leukemia Network 2022 classification, it is more of a static prognostic system. What does that mean? That we look at the mutations and the cytogenetics at the time of the presentation, and we classify patients into low risk, intermediate risk, and high risk, and based on certain mutations, sometimes we can add targeted therapy; however, the problem is that it sets a one-time like snapshot. You don't get to see what happens in the future. You cannot use information that happens in the future. Despite that AML is very dynamic, we have not been using those additional data points to help for prognostication and also help in making decisions.

Dale Shepard, MD, PhD: So, when you think about coming up with better ways to do this, this was a process that took a while. This is involved sort of multidisciplinary, in a way, approach, so tell us a little bit about how this all came together.

Moaath Mustafa Ali, MD, MPH: Actually, it started when I was a fellow at University of Maryland. I was attending a grand round, and Dr. Daniel Lobo, he's a computational biologist at University of Maryland. He presented, in a grand rounds, a nice presentation, how to use natural selection computational models to show interesting research. I thought it was very interesting, and I thought it would be relevant in acute myeloid leukemia. After he finished, I reached out to him. I set up a meeting, and we discussed that if we can do that in acute myeloid leukemia. When I was a fellow, I generated a database for acute myeloid leukemia. We started that from scratch. We collected data on almost 500 patients, and the data was unique because it was longitudinal. That means that we had data collected at different time points throughout the patient history, not just one data point set. Then, it went through, we probably took a year and a half to collect data, and then after we finished the data collection, I reached out to him again. We set up meetings. I helped them understand AML more, and he used his knowledge to help build this computational model or algorithm to predict AML. The data included patients from 2001, almost 14 years of data for AML patients, and the end date was 2021. Then, it took us almost three years to finish this study.

Dale Shepard, MD, PhD: Wow, so a labor of love and took a lot of time. I would see that presentation. I would say, that's awesome. That's really interesting. I might not have a clue how to start, and so oftentimes, I guess you don't know what you need to know to build these models, so how do you approach what patient factors, what disease factors, what treatment factors to collect and sort of think about incorporating into a model like this?

Moaath Mustafa Ali, MD, MPH: This modeling can be applied for AML or other diseases. The most important thing is you need to collect comprehensive amount of data that is accurate and in a longitudinal fashion. For example, you need to collect patient characteristics, their age, their ethnicity, smoking habits, alcohol intake habits, gender. Then, also, you need to collect characteristics on the cancer itself. For example, in AML, what is the type of the AML, the blast percentage, the mutations, the cytogenetic abnormalities? Then, also you need to collect data on the therapy, like what type of therapy were given, what were the blasts after you gave the therapy, and so on? Then, you also want to collect the data in a longitudinal fashion, so every time you get next generation sequencing, you need to collect that data, because all of that data enters this model, and that's what makes this model unique because it predicts outcomes in a longitudinal fashion, not just based on one time point at the beginning of therapy.

Dale Shepard, MD, PhD: So, complex mechanisms for a simple oncologist like myself, for hematologists, we hear about computational modeling, we talk about evolutionary computation. What does that mean?

Moaath Mustafa Ali, MD, MPH: Evolutionary computational modeling, it's a kind of computational algorithms that start with a number or a population of models that try to explain the relationship between data points or variables and an outcome. For example, it tries to understand the relationship between the cytogenetic abnormalities or specific mutations, like TP53 or MP-M1 or FLT3, and how is that related to the blast percentage? That was our outcome. Then, it looks to these models, looks to which one explains the data better, and looks to which one has the least amount of errors, and start excluding the ones that have high amount of error as if this is a natural selection process. Then, what happens, also crossing between these models… and then random alterations or random mutations to the model, also... This process is an iterative process that continues to compute over time to find a model that has the least amount of error. We repeated this 20 times, and we found in these models around 57 data inputs were important and explain these models and that's how it happened.

Dale Shepard, MD, PhD: Well, so certainly this has occurred over a long period of time. We have different new ways to do computations. AI comes into play. With some things we're doing, how much of a role could AI play in terms of redefining or helping define those models?

Moaath Mustafa Ali, MD, MPH: AI is very important, integral here. Now, how we can use AI to make this process easier. One process is to automate the data collection. For example, at Cleveland Clinic, we have generated this massive AML database. It has 1100 patients, and also a massive database of solid tumors, almost 6000 patients. Now I'm going to work with the AI team here at Cleveland Clinic, try to automate this, so we will use the data that was generated by a larger group of people and try to automate this to make this a faster process. So we'll be very unique at that step, hopefully ahead of other hospitals, actually. So, that's one point. Now, this type of evolutionary computational analysis is a kind of artificial intelligence analysis. There are different methods of AI or machine learning. There's something called ensemble or tree-based, such as random forest or gradient boosting. There's also the deep neural networks, which a lot of times need massive amount of data. Now, the study that we have done is explainable. It's not like sometimes there's a term people use, like black box AI, where you really don't understand the relationship between the variables and the outcome. Now, but this design we use, it's not like that. You actually see the data inputs, you see the data output, and then you can see the relation between them. For physicians and anybody who understands the science, will see the changes like real time and can confirm if it makes sense or not, so it's different than those type of machine learning. Also, what's nice about the method that we use, it can simulate data, generate these models using longitudinal data, but for example, random forest or gradient boosting cannot do that. Deep learning networks, like neural networks, they can, but they need extreme amount of very large databases.

Dale Shepard, MD, PhD: Yeah, so I guess the AI to extract data from the medical record is going to put fellows out of a job for...

Moaath Mustafa Ali, MD, MPH: At some point, it is possible, yes.

Dale Shepard, MD, PhD: Yeah. All of the database projects that come around.

Moaath Mustafa Ali, MD, MPH: Yes, yes, yes.

Dale Shepard, MD, PhD: So, developed a model. What were some of the more surprising findings?

Moaath Mustafa Ali, MD, MPH: So, a lot of things that we saw in the models are things that we expected. For example, like if we look to the occurrence of FLT3-ITD, we saw that the blasts are very high. If we saw, for example, certain cytogenetic abnormalities, we will see that the blasts are high or versus low. If you give certain kind of treatments, obviously the blast will go low, will decrease. So, those are things that we knew, but the significance of this is that it understood the relation between all of them at the same time. For a naked human eye, we can predict the relationship between one variable and another, maybe two variables together, maybe three variables together, but when you want to look into 60 variables at the same time, it's hard to understand the relationship between them. But at the same time, it's within the human limits to understand if you provide a model in front of them, so they can look into it and say, "Okay, yeah. It makes sense." So, understanding the hidden relation between different variables, that's something that was very interesting or something very interesting we found. For example, DNMT3A, depending on the context, sometimes can be associated with increased blast versus lower blast, so it shows the complexity of relationship between all these data points, which represents the reality, because we humans try to simplify things always, but sometimes we're not able to understand the very complex picture and relation between all of these variables at the same time.

Dale Shepard, MD, PhD: So, you had a really large number of data points. You talked about how you didn't do a black box concept for development, but really something that biologically made sense, but despite that, was there anything that sort of came in and stayed in the model as a factory like, "Wow. I would've had no idea that that would've been important"?

Moaath Mustafa Ali, MD, MPH: Most of the things that we included, they made sense, which is good, because at the end of the day, you don't want to see something that did not be completely unexpected. I think the complex relation, the two-step and the three-step relation was the most interesting, but there was nothing extraordinary, I would say.

Dale Shepard, MD, PhD: So, you've put together a model. What's the next step?

Moaath Mustafa Ali, MD, MPH: I think the leukemia community, as well as oncologists, we need to move to the next step. The next step is, instead of trying to prognosticate patients based on baseline factors, baseline variables, like the presence of this mutation only or the presence of this cytogenetic abnormality only, we need to start looking into the longitudinal picture. The method that we propose is one of the methods that we can use. Now, what I propose, and my hope is that this is going to happen sometime soon within the next several years, is that we need to collaborate together to generate larger set of data, so then the models can learn better and can give us more accurate estimates. That's very important, why? Because it can help us not only predict the outcomes and the survival, but it also can help us predict when the relapse may occur, and what is the best treatment choice if a relapse occurs? It can provide a lot of useful information. Unfortunately, at this time, the models that we use for prediction of outcomes, they are only limited in AML, in MDS, in ALL, all of these diseases, even in solid tumors, so this method can be, actually, applied to different cancers.

Dale Shepard, MD, PhD: At this point, if you think about the AML model, do you think it's going to have initially a bigger impact on when the patient comes in, how you treat them initially, or after they've completed their initial therapy, whether they go to transplant or those kinds of decisions, or both?

Moaath Mustafa Ali, MD, MPH: It will help make decisions at different stages of the treatment.

Dale Shepard, MD, PhD: And then, you mentioned some other diseases that this could be applied to. Are there any that you're currently working on?

Moaath Mustafa Ali, MD, MPH: For example, you can do a similar design in solid tumors. You can do it in acute lymphoblastic leukemia. You can do it in different diseases, but the thing here is that you need to collect the data in a longitudinal fashion. The treatments, what are the outcomes of each treatment when it was administered? You also need to collect a lot of genomic data. Genomic data is extremely useful. Now, in acute myeloid leukemia, because we tend to repeat the next generation sequencing a lot, we were able to do this such design, but if you know, eventually in solid tumors, this will become more and more standard of care, where you try to get mutation analysis, next generation sequencing at every stage of treatment and keep sampling. Obviously, the limitation here is the resources because these technologies are expensive, but they are necessary for accurate prognostication and also determining the best treatment.

Dale Shepard, MD, PhD: The point you just raised is it seems like, particularly solid tumors, there's a lot more variability in terms of when people get scanned and how often we get genomic data, so we make a genomic data diagnosis, and maybe way down the line when we're searching for other treatment options. So, do you think maybe these models could change the way we approach standard treatment, just so we can actually know a little bit more?

Moaath Mustafa Ali, MD, MPH: Definitely, because it's very important. It can help you, when you do the molecular testing, repeated testing, you actually can understand the clonal dynamics. For example, you're treating a patient with small cell lung cancer or non-small cell lung cancer. Then, after treatment and they have a residual disease, and then you do the molecular testing, you actually can see that maybe there's emergence of a new mutation, so that can help you make a decision quite early. My advice for all cancers, not necessarily in AML, but all cancers, my hope is that the molecular testing will become more frequent and done in a systematic way that can help us move the field further.

Dale Shepard, MD, PhD: Well, fascinating work. It certainly has taken you a long time to get here. It looks like it's been very fruitful though, and now that you have a sort of a framework, it looks like you're onto success in other areas, so appreciate you being here and give us your insights.

Moaath Mustafa Ali, MD, MPH: Thank you. Thank you for inviting me.

Dale Shepard, MD, PhD: To make a direct online referral to our Cancer Institute, complete our online cancer patient referral form by visiting clevelandclinic.org/cancerpatientreferrals. You will receive confirmation once the appointment is scheduled. 

This concludes this episode of Cancer Advances. For more podcast episodes, visit our website, clevelandclinic.org/canceradvancespodcast. Subscribe on Apple Podcasts, Spotify, or wherever you listen to podcasts. 

Thank you for listening. Please join us again soon.