Expert Q&A: Multi-Dimensional Analysis of Tumor Heterogeneity in CLL Enables Effective Therapeutic Selection


Chronic lymphocytic leukemia (CLL) is a blood and bone marrow cancer and the most common form of leukemia in adults. One of the clinical hallmarks of CLL is the high degree of variability in its disease course. Although targeted therapies for CLL exist, some patients harbor subpopulations of resistant leukemia cells that mediate disease recurrence. Understanding the heterogeneity of CLL tumor cells is therefore important for the development of novel treatment strategies.

Dr. David Andrews, Director and Senior Scientist in Biological Sciences at Sunnybrook Research Institute in Toronto, Canada, and colleagues have been using high-content image-based screening to address the current challenges in CLL and identify cohorts of patients whose cells behave in a similar way in response to treatment. In this article, David shares details of his recent studies and discusses his hopes for launching a powerful clinical trial in the future.

PerkinElmer: Why is CLL of interest to you? What are the current treatments and what challenges are posed by this disease?

Dr. Andrews: What stands out to me is that CLL remains a fatal disease. Patients might go into remission for extended periods of time, but they will not overcome it and we don’t know how to cure it. This means that everyone is ultimately resistant to treatment. A challenge is predicting how well patients will respond to treatment and for what duration.

In CLL there are two new targeted therapies: Ibrutinib, a Bruton’s tyrosine kinase inhibitor that disrupts signaling networks important for CLL cell growth and venetoclax, which inhibits Bcl-2 – a protein that prevents cell death and is overexpressed in virtually all CLL cells. It is currently difficult to know for any particular patient which drug is the better choice.

Moreover, although some patients respond to treatment, most harbor subpopulations of resistant leukemia cells that mediate disease recurrence. We are currently working on ways to identify subsets of patients for which we can predict the more effective treatment, and to aid drug discovery with the ultimate goal of finding better subset-specific treatment options.

PerkinElmer: How have you studied resistance to treatment in patients with CLL? Can you describe how you developed an in vitro model of the microenvironment?

Dr. Andrews: The problem with CLL is that many of the drugs used to treat cancer are designed to kill proliferating cells, but the cells in the proliferation centers of CLL patients are often highly drug resistant. Since we did not know the mechanisms behind this, we worked with clinicians to develop an in vitro model of the leukemic microenvironment.

We took circulating lymphocytes from CLL patients and added to the growth media both interleukin 2 (IL2) and resiquimod, a medium we call 2S. The cells then behaved like the tumor driver cells in the lymph nodes or bone marrow, proliferated, and became very drug resistant. Media like 2S are not a perfect model – if we overdo it, the cells hyperproliferate and die in vitro, whereas in the body they won’t – but it gave us a first step.

We then demonstrated that most circulating cells from CLL patients die after adding just 10 nanomolar of venetoclax. However, most of the cells became resistant when we used 2S media, even at 100 nanomolar of venetoclax. For our first paper we looked at these findings and asked: Is this something we have told the cells to do or is this inherent in a subset? Our rationale was that if 2S provides an instruction set, we should be able to find kinase inhibitors that block 2S induced signaling and cells should once again become sensitive to venetoclax.

PerkinElmer: How did you use high-content image-based screening to identify kinase inhibitors that overcome venetoclax resistance?

Dr. Andrews: We screened a kinase inhibitor library of over 300 compounds in combination with a very low dose (10 nanomolar) of venetoclax to find combinations that would dramatically increase cell killing for a significant number of patients. However, finding a high kill can be problematic because cell death is stochastic; at any one time not all the cells are dying.

You need to be able to identify cells that are just starting to die, those that are in the middle of dying, and those that have already passed the point of no return. If you’re only looking at the ones you’ve killed, you are missing a lot of information and you can’t just wait longer or the cells that died early will have fragmented and will no longer be visible. This is where high-content screening comes in. We generated a multiparametric descriptor that when used with high-content imaging resulted in a positive correlation between outcome and a snapshot of the patient cells at a specific time.

The screen identified sunitinib, a broad-specificity kinase inhibitor, as the most effective compound sufficient to counteract 2S media-induced venetoclax resistance in most patient cells. We further discovered that the mechanism of resistance for a large subset of patients was upregulation of two other inhibitors of cell death, Bcl-xl and Mcl-1.

This is important because the whole idea of personalized medicine is to find a treatment specific for an individual. However, a treatment will not be approved unless you have a significant number of patients that behave the same way in a clinical trial. So, identifying a subset is ideal. What surprised us in this study was that we could only use pan kinase inhibitors to shut down the upregulation of both Bcl-xl and Mcl-1, and that was disappointing as we were hoping for something more specific.

For our latest approach we took patient cells, put them into the 2S media plus IL4, and identified at the outset six different patient types, or cohorts, just from the images of otherwise untreated cells. It appears that those images are then predictive of what the drug responses are going to be. By identifying these cohorts, we hope that we have solved the problem that we had with the original study relating to pan kinases inhibitors. We are now identifying more specific treatments for each cohort.

PerkinElmer: How are you testing combination therapies for these cohorts? How does high-content screening help you achieve this?

Dr. Andrews: We performed this in a 1536-well format, where CLL cells from each patient are seeded onto one plate, and everything is replicated. We can test 750 different combinations of one, two, or three drugs for each patient and can measure hundreds of image features per cell. It is an enormous amount of data.

One issue with high-content screening is the curse of dimensionality – if you have a relatively small number of patients compared to the number of features measured by high-content screening you always worry about overtraining. To address this, we took all the true features and reduced them to just six biological features.

For example, we might have a set of between 120 and 135 true features from the images that together represent a measure of stress, or quiescence, or apoptosis. These are the biological features, and so by doing this we can reduce the true features to six biological features in such a way that they retain biological meaning.

Once we could identify these individual cohorts of patients, we found that each cohort had a different response to treatments, and that patients within the same cohort responded similarly. This means that it may be possible to use the same treatment combination for all the patients in a cohort and this gives us the potential opportunity of conducting clinical trials. We're starting to work on possible trial designs with clinicians so that we can test these hypotheses.

PerkinElmer: Did you look at the genotype of these patients? Was this affecting patient response?

Dr. Andrews: Yes, and we found that the genotype is not predictive of how patients are going to respond to drugs. In our first study, we looked at the genotyping and methylation patterns and we saw a beautiful correlation with disease severity and zero correlation with drug response.

PerkinElmer: What challenges did you face in your studies and how did you address these?

Dr. Andrews: The most challenging issue for us was dye toxicity. If you add reagents to cells that are themselves toxic, then you will get a response that is convoluted between the toxicity of the drugs you are trying to assay, and the toxicity of the dyes you are using to make the measurements with. This is problematic when you are looking at cells because the time window becomes very narrow if you are using dyes that are influencing the outcome. It means that all cells have to be imaged on time; if they are imaged too early or late, the convolution of data with the toxicity from the dyes will be different and your answer will be smeared out and hard to interpret.

In our case, we used two dyes that were reasonably non-toxic and one that was moderately toxic. This meant we had to have a single timepoint, which was unfortunate. We had to work very carefully to avoid batch effects. We have subsequently made our own dyes that are non-toxic to cells to remove the convolution problem.

PerkinElmer: Were there any notable surprises in your studies?

Dr. Andrews: One surprise for us was that we sometimes got responses that seemed counterintuitive within a cohort. For example, you might get a cohort of patients that, in response to a particular set of drugs, will have a large increase in cells that die but there will also be a large increase in ones that exhibit a phenotype where they hunker down and wait it out. It is not clear why that would be at the moment and we are currently trying to study this.

PerkinElmer: What are your next steps and hopes for the future?

Dr. Andrews: For CLL, we are hoping to get to the point where we can launch a powerful clinical trial, but this has some challenges. Up until now, we have conducted several small trials and tried to compare results. We need to accumulate enough pre-clinical data where we can match the patient response to our prediction. With CLL this can be difficult because it is a chronic disease and many of the CLL cells that we have imaged are from patients that may not need treatment for another five years or more.

We have also adapted our method to locally advanced breast cancer and ovarian cancer. We have recently been funded for a trial in ovarian cancer, where we will take samples from the patients and make a prediction as to how they will respond to treatment to see if we can identify the subset of patients that will respond. Again, it is a small trial and hopefully we will have enough patients enrolled for it to be powered. We are also starting work on endometriosis and keloid scarring; it is exciting to try to extend this approach to other diseases.

Dr. David Andrews, Director and Senior Scientist in Biological Sciences at Sunnybrook Research Institute, Toronto, Canada

Further reading

Oppermann S, Ylanko J, Shi Y, Hariharan S, Oakes CC, Brauer PM, Zúñiga-Pflücker JC, Leber B, Spaner DE, Andrews DW. High-content screening identifies kinase inhibitors that overcome venetoclax resistance in activated CLL cells. Blood. 2016 Aug 18;128(7):934-47. https://ashpublications.org/blood/article/128/7/934/35889/High-content-screening-identifies-kinase