Harnessing Machine Learning for Mantle Cell Lymphoma Risk Prognostication With Adrian Mosquera Orgueira, MD, PhD
In this interview from the 2023 American Society of Hematology (ASH) Annual Meeting, Oncology Data Advisor speaks with Dr. Adrian Mosquera Orguiera, Clinical Hematologist at the University Hospital of Santiago de Compostela, about his research team's novel machine learning model that incorporates patient characteristics and histological subtype into risk prognostication for mantle cell lymphoma (MCL). In addition, Dr. Mosquera shares more of his research utilizing artificial intelligence for designing more risk-tailored therapeutic approaches and facilitate personalized medicine in hematologic malignancies.
Oncology Data Advisor: Welcome to Oncology Data Advisor. Today, we're here at the ASH Annual Meeting, and I'm joined by Dr. Adrian Mosquera. Thank you so much for stopping by today.
Adrian Mosquera Orgueira, MD, PhD: Oh, it's my pleasure to be here with you. Thank you.
Oncology Data Advisor: To start off, would you like to introduce yourself and tell us a little bit about what you do in your work?
Dr. Mosquera Orgueira: Yes, I'm a Clinical Hematologist in Spain. I devote most of my time to the treatment of patients with lymphoma and chronic lymphocytic leukemia (CLL), but I'm also specialized in the application of genomic and flow cytometry in clinical decision making. Through these data-driven approaches, we ended up developing of a research team, which is fully focused on artificial intelligence (AI) for different applications in clinical decision making through the integration of imaging, flow cytometry, genomics, and clinical data into the modeling strategy of clinical outcomes. We are very interested in data-driven hematology, and this involves a lot of approaches that use AI, particularly machine learning.
Oncology Data Advisor: Awesome. So, we're talking today about your study on harnessing machine learning and histological subtype for enhanced MCL prognostication of survival. For a little bit of background about this, would you like to tell us about the MIPI, the Mantle Cell Lymphoma International Prognostic Index? How is it currently used in MCL, and what limitations does it have?
Dr. Mosquera Orgueira: The MIPI score is actually a great score, because it is a quantitative progression–based model, and it provides patient-level statistics. It's been one of the few places in hematology where a good score has been developed in this way. This score obviously, at this moment, is not being used for risk-adapted therapy. In the future, maybe with the advent of Bruton tyrosine kinase (BTK) inhibitors upfront, we may be more persuaded to use these kinds of scores to do a treatment-adaptive or risk-adaptive strategy for MCL patients. This score can also be used to cluster patients into different groups of risk, as many scores do.
But the greatest advantage, actually, is that you can also drive patient-level statistics. Obviously, the limitation is that it doesn't incorporate, for example, histological data. We know that histological subtype is very important. It does not incorporate other molecular markers of high risk, such as tumor protein 53 (TP53). Overall, no score is definite. You just need to keep improving the scores over time, with additional data and with more contextual data. As the scores are developed and new treatments appear, the scores lose some value. You need to retrain the scores towards a new context of therapeutic approaches available in clinical practice.
Oncology Data Advisor: That's a great overview, and it's very helpful for understanding this. I'm excited to learn more about the model you and your team developed. How did you decide to approach exploring these limitations via machine learning?
Dr. Mosquera Orgueira: This was a combined effort by different hospitals in Spain that gather information from MCL patients, and then from a hospital in Taiwan that shared data for external validation in a very different clinical context. Our strategy was to evaluate if we could further refine the MIPI score—not change it but refine it with more information. In this case, we ended up with a model that, aside from the MIPI groups, included age and histological subtype as an independent part of the prognostication. We integrated it through a machine learning analysis. The benefit of this is that we ended up with quantitative scores that can be used at the patient level, not in clusters or at the group level.
This is interesting because it helps us refine groups of patients who are actually of higher or lower risk. Obviously, in the future, we may need to add some other information—for example, TP53 mutation, which we know is important in MCL, and also histological variables or gene expression signatures. But the more complex you make the model, the less applicable it is. In this case, using this very simple data, which is based almost on standard diagnostic routines, makes this kind of approach more widely usable, even in lower-income, middle-income, and also high-income countries.
Oncology Data Advisor: That's fascinating that it can be used in all those different income settings.
Dr. Mosquera Orgueira: Yes, I have the feeling that we tend to do medicine for high-income countries, because we are biased towards our imminent environment. But we need to think that there are lots of countries where most of the human population is living, and they need usable models for their patients. Actually, we collaborate a lot with Latin American countries in different projects, not just in MCL, but other disorders too. They like to become part of the validation and even training of these models in their clinical context. This is important, because at the end of the day, what you want is to benefit as many patients as possible and make treatments affordable by being capable of directing more expensive therapies to patients who really need it.
Oncology Data Advisor: Absolutely. It's not only about developing these technologies, but also making them accessible to everybody.
Dr. Mosquera Orgueira: Yes. Every technological advance has, as a drawback, inequality in its access. One way that machine learning can help is by reducing this inequality through creating systems for advanced diagnostics, prognostication, and drug response prediction, which can help us do personalized or precision medicine without the need of very complex data.
Oncology Data Advisor: Absolutely. So, how effective did the model prove to be, and how is it working so far?
Dr. Mosquera Orgueira: The model has been validated in another Spanish hospital cohort and in a Taiwanese cohort, and we actually observed high accuracy, which was superior to the MIPI score. This is normal, because we are incorporating another variable, which is the histological subtype. This variable actually adds independent prognostic value.
Now, another group from our collaborating environment has also used the same data to create another model for progression of disease within 24 months (POD24) prediction. These patients who have early relapse, we know they are of very high risk, but the problem is that you only know the patient is of high risk while he or she relapses, and the idea is to predict that. The same strategy with a very similar model and a few variables, which also include histological subtype, has been presented here at ASH. It's promising to risk stratify patients in early relapse and late relapse, and maybe do clinical trials with more intensive approaches in patients who are actually of very high risk.
Oncology Data Advisor: Great. What are your next steps that are planned for the model?
Dr. Mosquera Orgueira: The next steps are to refine it, validate it in newer cohorts of patients, and test how it behaves in different clinical trials. We are involved, as many other researchers are, in clinical trials that try to get rid of autologous stem cell transplantation or even reduce chemotherapy as much as possible by adding venetoclax, anti-CD20 antibodies, or BTK inhibitors combined in the backbone of treatment. This will probably need a refinement of the models, because once you change the treatment with different molecules, then the risk will change. You need to adapt models constantly to that. This has happened in many disorders.
We also know that patients might be high-risk for one therapy but standard- or low-risk for another therapy. What would be nice is to be capable of entering or navigating through the data of these advanced clinical trials. I don't know if that's going to be possible, but making predictors with those data would be great. Actually, we are also very interested now using real-world data that we have at our institution to apply machine learning to histological data. That's not historical human-based classifications, but the raw image of the biopsy, the raw image of the positron emission tomography/computed tomography (PET/CT), and the molecular signatures based on gene expression profiling that we have developed at our institution. We can maybe improve the prognostication, particularly by the biological perspective of the disease.
For these patients, sometimes the risk that we model is not just related to the disease, but also to the global condition of the patient. Their age, their comorbidities, et cetera, impact mortality. It's also important to know to what extent a patient is likely to die early due to the lymphoma, not due to other comorbidities that he or she may have. This is refining the biological compartment and giving a high-risk biological signature could be of the utmost value for MCL. These are some of the approaches. For example, the MCL35 gene expression signatory is over there. The problem for MCL is that it's difficult to do molecular tests that are available, particularly for this condition, because it's quite rare. In our case, we've developed a signature for diffuse large B-cell lymphoma (DLBCL), which is a widely diagnosed lymphoma, and we observed that it is also predictive of response in mantle cell lymphoma. We are working on standardization of this test so that it may be used in clinical practice in the coming years.
Oncology Data Advisor: This is fascinating to hear about, and it'll be so exciting to see all the ways these avenues are explored in the future. Is there anything else you'd like to mention about your research?
Dr. Mosquera Orgueira: Yes, actually. We are very concerned about the need for data-driven approaches in medicine—producing our decisions based on experience and augmenting our capacity of predicting outcomes and refining treatments through data-driven approaches. As such, we are very happy to work in many projects with colleagues from Europe and the United States and Asia.
I think that there is a momentum for AI in medicine. There is a growing interest in guiding treatments by risk profiles. Pharmaceutical companies are becoming more and more aware of this. I think it's helping us to get rid of the idea of "one size fits all" for cancer treatment. Rather, we are moving towards a more precision medicine approach, which will need a lot of AI, obviously. We are fully committed to helping in this transformation. We are now working on very insightful projects, for example, in myeloma. We hope that in the coming three or four years, we'll have advanced tools available, based on artificial intelligence, to guide therapies in hematological tumors.
Oncology Data Advisor: Amazing, this has been so fascinating to learn about. Thank you so much for stopping by today to tell us all about your research.
Dr. Mosquera Orgueira: You're welcome. It's been a pleasure.
About Dr. Mosquera Orgueira
Adrian Mosquera Orgueira, MD, PhD, is a Clinical Hematologist and Lead Researcher of the Computational and Genomic Hematology Group at the University Hospital of Santiago de Compostela in Galicia, Spain. His research focuses on the application of genomics and artificial intelligence in precision medicine.
For More Information
Orgueira AM, Carbita A, Garces VN, et al (2023). Beyond MIPI: harnessing machine learning and histological subtype for advanced MCL prognostication of survival. Presented at: 2023 ASH Annual Meeting. Abstract 4419. Available at: https://ash.confex.com/ash/2023/webprogram/Paper179174.html
Touzon AC, Orgueira AM, Garces VN, et al (2023). A novel machine learning model to predict early relapse in mantle cell lymphoma (MCL). Presented at: 2023 ASH Annual Meeting. Abstract 1675. Available at: https://ash.confex.com/ash/2023/webprogram/Paper181628.html
Transcript edited for clarity. Any views expressed above are the speaker's own and do not necessarily reflect those of Oncology Data Advisor.