Predicting Response to Immunotherapy With a Mathematical Model

Matthew D. Hellmann, MD

Benjamin Greenbaum, PhD

Immunotherapies such as immune checkpoint inhibitors targeting the programmed death 1 (PD-1) and cytotoxic T-lymphocyte–associated antigen 4 (CTLA-4) pathways have the ability to significantly extend survival for patients who respond to these treatments. The problem is that only a proportion, and in the case of many tumor types, only a minor proportion of patients achieve durable responses and it is difficult to find out who is likely to respond. Expression of the programmed death ligand 1 (PD-L1) protein on tumor cells is one marker that clinicians have used to predict response to these immunotherapies, but patients whose tumors don’t express the marker can also respond to treatment.

Researchers at the Icahn School of Medicine at Mount Sinai in New York City have developed the first mathematical model that they say can predict whether or not a patient will benefit from checkpoint inhibitor immunotherapy. Their work was recently published as two papers in the journal Nature.[1,2] We are joined by two of the study’s authors, Benjamin Greenbaum, PhD, an assistant professor at the Tisch Cancer Institute at the Icahn School of Medicine, whose team developed this predictive model, and Matthew D. Hellmann, MD, a medical oncologist who specializes in the care of patients with lung cancer at Memorial Sloan Kettering Cancer Center in New York City.

-Interviewed by Anna Azvolinsky

Cancer Network: Dr. Hellmann, could you lay out the problem of predicting response to immune checkpoint inhibitors and factors that are known to be associated with response to these therapies?

Dr. Hellmann: Particularly in the last 5 years, one of the major advances in cancer care as a whole has been the advent of immune checkpoint inhibitors as a therapy for patients with a variety of cancers. It’s been more than a century of observations and experiments trying to identify ways in which we can leverage our own body’s immune system to fight cancer, and there have been a few successes along the way. Since 2012 onward there has been a major explosion of drugs targeting PD-1 or PD-L1, which has led to indications for several different therapies targeting this pathway in multiple cancers; they have proven to improve survival for a subset of patients with a variety of cancers such as melanoma, lung cancer, kidney cancer, Hodgkin lymphoma, and mismatch repair–deficient cancers, just to name a few. That progress has been very profound, it’s happened quickly, and it is a tremendous opportunity for patients. However, there have been some limitations to the therapy that require a whole lot of effort as we move forward, particularly in trying to better understand why some people benefit and other people don’t benefit at all.

Thinking about lung cancer as an example, probably about 20% of people treated with these therapies respond, and when they respond, they can respond both deeply and durably, lasting much longer than with other cancer treatments. But, it also means that about 80% of people are not benefiting at all. So, understanding why that is, why there are such different outcomes for some people compared with others and being able to predict who is who beforehand could make a huge difference in terms of our ability to use these therapies effectively and be able to identify more rational combinations for the future. It’s been phenomenal, and yet there is a lot of opportunities to improve.

Work that has been done previously has highlighted, in lung cancer, melanoma, bladder cancer, and microsatellite instability–high cancers, that there seems to be an association between the number of mutations in a given cancer and the likelihood of benefit, and that has at least given some initial insights in trying to better predict who the people are who are likely to benefit. An open question has been why more mutations are associated with a better benefit; a leading hypothesis is that each time a mutation is made it has the possibility of creating a new antigen that is only seen in the tumor and that may allow and direct an immune response to the cancer. So, that at least in a crude way, the more mutations you have, the more likely you are to have a good antigen that allows the immune system to recognize the cancer. This has prompted a lot of work to try and understand the ways in which a given mutation yields an effective antigen to predict what tumor antigens might be present.

Experts are also now trying to develop cancer vaccines based on predicting the right antigens that might be in a given tumor, but the models that have been used to do that thus far have not been terribly effective. They have not been able to zero in on the specific neoantigens that are relevant for any given patient. In the experiments done so far, if you were to predict, say, 100 new antigens that might matter, maybe only one of those is experimentally validated, so there is a lot more room for improvement for more nuanced approaches in thinking about the neoantigens that matter in a given tumor. The paper done by Łuksza et al is a real step forward in providing new insights into new mechanisms to understand what the specific antigens are that might matter in a given cancer and how to identify them from a patient’s tumor sequence.

Cancer Network: Professor Greenbaum, can you describe the mathematical model that you and your colleagues created to potentially predict response to these immunotherapies?

Professor Greenbaum: Sure. Our goal was to try to create a mathematical framework that incorporated exactly the features that were just described into a model-namely, what is the likelihood that one of these novel cancer antigens created through a mutation in the tumor will be presented and recognized by the immune system? And how are those antigens distributed over the clonal structure of the tumor? So, to integrate all of these things into one mathematical model and to estimate how that would affect the evolution of the tumor under therapy, we tried to make a model as simply as possible using the current data available that reflected that information, and then as a sort of benchmark, to see if that model improved upon the predictive power associated with the correlates that Dr. Hellmann just described.

Cancer Network: Can either of you talk about how you and your colleagues tested the model on existing cancer patient data and what the results were?

Professor Greenbaum: There were three published studies that established that the mutational burden of the tumor correlated with response to some degree. A model like this has to be trained, there have to be some parameters that must be inferred from the data, and so we used one of those datasets to train the model and then we tested it on another dataset, and then we reversed the process of training it on one dataset and testing it on another, since this was not done on a prospective study. That was the general approach to see if we could get a consistent model across all three of the previously published datasets that we were able to use for the study.

Dr. Hellmann: One of the things that I thought was particularly interesting about your approach was some of the lessons learned from thinking about antigens in viruses and particularly in influenza, and how that information has been used to think about building a flu vaccine, and I wondered if you had a way of thinking about that and if you wanted to describe how you’ve applied things like that in the context of cancer.

Professor Greenbaum: That is a really good point. The inspiration for trying this approach was that I have a lot of experience in virus evolution and influenza and it’s a system that I am particularly interested in. Dr. Łuksza, the first author of this paper, in work that I was not involved in, had shown that taking into account the phylogenetic structure of the virus evolution and the antigenicity of viral antigens allowed one to improve upon predictions for what would be next year’s circulating flu strain, and that one could incorporate those predictions into a vaccine prediction pipeline for influenza. It was the success of that approach combined with the emerging hypothesis that these therapies were working in high mutational burden tumors, that they were being driven by the recognition of neoantigens across a tumor’s clonal structure, that inspired us to give this a try here. I think that it’s one of the novel aspects of this approach.

Cancer Network: Professor Greenbaum, what’s next as far as this model? Is there anything missing from the model that you plan to add based on existing data to improve it?

Professor Greenbaum: I think there is a lot of room for improvement. I think that we all view this as a proof of concept that this kind of model that takes into account underlying processes offers the possibility for better predictive values in the future. But all of those steps in the underlying process associated with the processing and presentation of antigens in the tumor have a lot of room for improvement. Information about the environment that the tumor is growing in-in terms of how it’s interacting with the immune system-is one of the things I liked about this, that it provides a framework for looking at those problems, and as one’s ability to assess those things improves, one can improve on this approach by incorporating that new information.

Cancer Network: Lastly, Dr. Hellmann, do you see this version or a future version of this model as being used in the clinic?

Dr. Hellmann: I do. I think there are three main ways that I can imagine this being applicable sometime in the near term. The first is trying to help answer this key question of how do we identify patients who are most likely to benefit? The capacity of this model to integrate both the active antigens in the tumor and other features of the tumor, including inflammation of the tumor, mutation burden of the tumor, and other clinical features, will all help us to develop a more precise predictive tool to apply in the future. So, simply as a predictive tool, it can be very powerful and a component of more personalization in the use of immunotherapy as we move forward.

I think another scenario in which I can imagine this model being helpful is in studying biology. I think there is lot of effort ongoing to try to understand what the effective antigens are in a given tumor, and if this model gives us a more efficient pathway towards understanding how a given tumor interacts with the immune system, then what are the key features that determine an effective antitumor immune response? I think that this could be a very powerful tool and one that builds upon what is currently being done, but allows things to be done in a more efficient and nuanced way. A third way that I can imagine this being applied is in the context of multiple companies developing vaccine approaches to try to study the genetics of a tumor and try to build a vaccine based on the genetics of an individual tumor. There are a variety of companies that are making peptide vaccines and RNA vaccines based on studying individual tumors, but integrating this additional piece of information and increasing the accuracy of understanding what the right antigens are could yield more effective vaccines. I can imagine all of this being possible in the near term.

Cancer Network:Thank you both for joining us today.

Dr. Hellmann: Thank you.

Professor Greenbaum: Thank you.

References:

1. Łuksza M, Riaz N, Makarov V, et al. A neoantigen fitness model predicts tumour response to checkpoint blockade immunotherapy. Nature. 2017;551:517-20.

2. Balachandran VP, Łuksza M, Zhao JN, et al. Identification of unique neoantigen qualities in long-term survivors of pancreatic cancer. Nature. 2017;551:512-16.