Nucleosome Patterns in Circulating Tumor DNA Reveal Transcriptional Regulation of Advanced Prostate Cancer Phenotypes - Gavin Ha

June 8, 2023

Gavin Ha and Andrea Miyahira discuss Ha's research into the transcriptional regulation of advanced prostate cancer phenotypes. Dr. Ha's team uses circulating tumor DNA (ctDNA) in liquid biopsies to overcome challenges with traditional tumor biopsies in metastatic cancer patients. The focus of the research is on tumor phenotype, which may reveal why some tumors resist certain therapies. A new tool called Griffin, developed by the team, can profile nucleosomes from ctDNA. The team also developed machine learning models to classify and predict dominant and mixed prostate cancer phenotypes. The findings indicate potential for using preclinical models and ctDNA to further research and identify predictive biomarkers for treatment outcomes. Future steps involve expanding studies to assess the predictive power of these tools in therapy responses.


Gavin Ha, PhD, Assistant Professor, Herbold Computational Biology Program, Public Health Sciences Division, Assistant Professor, Human Biology Division, Fred Hutch

Andrea K Miyahira, PhD, Director of Global Research & Scientific Communications, The Prostate Cancer Foundation

Read the Full Video Transcript

Andrea Miyahira: Welcome. I'm Andrea Miyahira and I'm the senior director of Global Research and Scientific Communications at the Prostate Cancer Foundation. Today I'm joined by Dr. Gavin Ha, an assistant professor at Fred Hutchinson Cancer Center and the PCF 2019 Young Investigator. Dr. Ha's group has recently published the paper Nucleosome Patterns in Circulating Tumor DNA Reveal Transcriptional Regulation of Advanced Prostate Cancer Phenotypes in Cancer Discovery, which we will be discussing today. Dr. Ha, thank you for joining us.

Gavin Ha: Thanks so much. It's great to be here. I'm very happy to tell you about more of our work. Liquid biopsies is certainly an area of great interest to us and we think in the field of prostate cancer and other cancers as well. Here I'll tell you a bit about how we've innovated and developed new methods to really capture signals that were previously not captured from liquid biopsies.
To kind of set the stage for this as motivation, tumor biopsies are of often quite difficult to access, especially in metastatic cancers and particularly prostate cancer as well. Many issues related to that with comorbidity, invasiveness, the tissue quality may not be of sufficient use, and also, single biopsy may not represent the entire disease within the body. And, of course, then for monitoring applications, repeated biopsies are usually not feasible. And so, what we reasoned is that using circulating tumor DNA will allow us to open more doors and opportunities to be able to tackle a question in precision oncology. One, for example, tumor classification. Understanding how the tumor has changed, and then also coming up with better ways to see if this can inform treatment selection and planning.

Current methods in the clinic and in research are quite focused on genotyping circulating tumor DNA. That is, for example, looking at genomic abnormalities or mutations to be able to monitor the tumor burden directly from the blood. We can also look for resistance mutations. But it turns out that sometimes for certain patients, their tumors that are resistant to therapy may not be fully explained by genomic alterations. And so what we wanted to really get at now is can we better get a better sense of tumor phenotype directly from ctDNA, where the phenotype is what we want to be using to represent or to be able to capture additional information that may tell us more about why patients' tumors are resistant to certain therapies.

In prostate cancer, we know of this very well established transdifferentiation model where adenocarcinoma, or AR-positive prostate cancer, can transdifferentiate in the selective pressures during treatment with androgen receptor pathway inhibitors, or signal inhibitors, into a neuroendocrine or small cell-like prostate cancer we call NEPC. And so, this phenotype switch does have some genomic alterations that are associated with it, but it doesn't always fully explain every single tumor. Some of these alterations can also be present, adenocarcinoma that hasn't transitioned as well. And so we want to be able to better capture this information directly from cell-free DNA.

Cell-free DNA are fragments that we're able to capture from the blood, and in particular we look at the plasma, and their characteristic 167 base pairs because they're protected by nucleosomes. That is an important observation, because when we take these fragments and we look in the genome of where they map, we're also getting a footprint of where the nucleosomes are found. Why is that important? Well, we know that the nucleosome positioning within a cell is a consequence of transcription regulation. For example, where genes are actively transcribed, nucleosomes will vacate so that transcriptional machinery can combined as well as transcription factors and other factors that would bind to initiate transcriptional programming.

And so, we've developed this tool called Griffin, which essentially profiles nucleosomes from circulating tumor DNA, cell-free DNA, and we're able to map all these reads and apply the right normalization approaches so that we can capture these signals from the various different cells that are shedding tumor DNA. The question here is, can we use Griffin to study the transcription regulation in prostate cancer phenotypes?

We were faced with one challenge initially, and of course many other challenges along the way, but the first one was, using patient plasma ctDNA tends to be either lower and undetectable for many patients, some patients have high, and it really reduced the number of samples that we're able to appropriately classify or build our models in development for detecting these phenotypes. And so, we teamed up with Dr. Peter Nelson at the Fred Hutch and also colleagues at University of Washington and Navonil De Sarkar who is now at the Medical College of Wisconsin.

We came up with this idea to look in preclinical models at patient-derived xenografts. We reasoned that if we are able to collect plasma from these mice, because the tumors are engrafted from human tumors, the human tumors are engrafted in the mouse, if we're able to isolate cell-free DNA from these mice and we do bioinformatic subtraction of any mouse-related contamination, we would have pure human ctDNA for us to do our methods' development. It turns out that worked very nicely for us and we're able to apply this to 24 different lines, different models of PDXs, including adenocarcinoma as well as neuroendocrine, and also a couple lines with AR low PC.

Importantly though, these PDX models also have the tumor tissue that are well characterized, and here we perform additional characterization of it for chromatin and histone marks. There we have this ground truth that we are able to compare all of our results from cell-free DNA. So it really provides a great resource, but a nice framework to be able to do all of our methods' development.
The first thing we did was we looked, using Griffin, at 338 transcription factors. We looked at all the binding sites of these transcription factors that would then infer, from our tools, the activity of these transcription factor binding. And we see very clearly that there are known transcription factors that are active in adenocarcinoma, PDX lines, and in the neuroendocrine PC lines, we see factors that we also know that are active, for example, ASCL1. So we know that the transcription regulation inferred by Griffin for these transcription factors are active as we expect.

And so then we developed machine learning models so that we can apply what we learned from the PDX models directly to patients, and in particular here, we want to be able to classify adenocarcinoma and neuroendocrine. And so here on the left, just showing you how we can, as a schematic, take what we learn from the plasma ctDNA from these PDX models into two models that we developed. And now we're analyzing clinical patient specimens. For the first model, which we call ctdPheno for classification, we applied this to a cohort of 101 patients from Dana-Farber Cancer Institute that was published in 2021 in Clinical Cancer research. Interestingly, this data is what we also call low-pass whole genome sequencing, which is very cost-effective. From this data, we're able to classify, with an area under the curve of 0.96, as being neuroendocrine or adenocarcinoma. At an optimal cutoff score, we have a sensitivity of 90.4%, specificity of 97.5%.

But we also know that during transdifferentiation, it's possible to have mixture of adenocarcinoma and neuroendocrine cells at the same time. And so, we also wondered whether we can develop methods to be able to capture the mixture that's present and then quantify what that mixture would be. So we came up with this tool called Keraon, which then we applied to another cohort of 47 samples from 27 patients to be able to classify, first, dominant phenotype of ARPC in the blue, as well as dominant any PC phenotype, but also for a subset of patients who, from their clinical histories, have, and also for some we have biopsies, evidence of mixed phenotypes. We see that Keraon performs quite well for the dominant phenotypes with the 97% accuracy. And then for the mixed phenotypes, we have an 87% accuracy of calling the presence of both phenotypes, adenocarcinoma and neuroendocrine.

In summary, we think that using preclinical models, PDX mouse plasma, really is a unique resource and we're able to build this out, and now we're happy to also share this sequencing data with the research community as well. There's many other things that we can also look in here and we have active research continuing on that front. I showed you a little bit of it as well, transcriptional activity that we can also tease out from cell-free DNA, mostly here from the PDX models, again, because it's a platform really for us to develop new methods to really dissect the specific gene activity that we can look at. Here are listed some of the factors that we were able to ascertain from cell-free DNA in this manuscript. We also develop new classification methods that can then predict dominant and mixed CRPC phenotypes directly from patient plasma. And of course, future work here, we would take these models and apply it to additional studies with larger cohorts that we can identify potential predictive biomarkers, treatment outcomes.
And so, obviously a lot of individuals and labs here to thank. Of course, Robert Patton, who developed a lot of the computational methods, Navonil De Sarkar who spearheaded the PDX studies in Peter Nelson's lab, and many collaborators from the University of Washington, Dana-Farber Cancer Institute, and Fred Hutch. Also our funding sources. I was like, not Andrea said at PCF YI in class 2019, Navonil was also PC YI in 2019, so we're very grateful for the support from PCF and all my other funding sources as well.

Andrea Miyahira: Thank you, Dr. Ha. This was a very comprehensive multi-omic ctDNA profiling study, and it's interesting to see what the pure circulating tumor DNA looks like when it's unmarked by DNA from non-tumor cells. What were the most surprising findings made using this model and what have we previously missed?

Gavin Ha: Yeah, those are great questions. I think the surprising thing for us is how much information we can actually get from these PDX models. It's a preclinical model that just continues to give. We've learned so much from them and many, hundreds, of labs throughout the world have been using these, and we're very lucky to have most of these resources available to us here in Seattle.

Yeah, I would say that we're really expanded on what we can do with ctDNA, especially with this resource. Think about how the blood is normally just thrown away. And so, really, it's the insights from Dr. Nelson and others to start collecting it and see if we can use that for development. The other surprising finding that I think previously may be underappreciated, let's say, is really addressing question of phenotype heterogeneity.

From imaging analysis and perhaps from genomic clonality analysis in treatment evolution, yes, you can find those as well, but they really kind of get at either the throughput issues with imaging, but also focusing only the genomics with tissue biopsies. Here we're really tackling a question I think is important, because we know many patients have multiple phenotypes within their body, and it could be either within a tumor or across different lesions within the body. And so I feel like even though it's a proof of concept right now, we're very excited about being able to address that question.

Andrea Miyahira: Can you compare and contrast what you've done here with other ctDNA assays that are currently being used clinically?

Gavin Ha: Right. Yeah. I think from my knowledge of what assays are out there, there is mostly a focus on looking at genotype or genetic abnormalities. I think that has great clinical utility. Certainly there are some clinical decisions that can be made from that therapeutic choices. But epigenetic profiling is also used for cancer, but has been really focused on cancer detection right now in the field and maybe perhaps for residual disease detection, but less so for doing phenotyping for therapy selection. And again, we know that it is really the phenotype of the cell that is ultimately what is going to be either resistant to therapy or responsive to therapies. That might not be fully explained by the genetic alterations.

What we're also excited about here is that we can leverage this low-cost approach to be able to take advantage of standard sequencing platforms. So, really no specialized assays here. And so I think there's great potential for adding to the current clinical landscape of these different tests.

Andrea Miyahira: Well, what are your next steps for translating your tools and methods into clinical applications for patients and what clinical uses are you envisioning?

Gavin Ha: Yeah, lots more work to do. I think we obviously would like to expand all these studies to really test whether our tools can predict patient response to specific therapies. I think we have a toolkit now to be able to refine it and tune it so that it can be potentially a companion diagnostic for specific therapies, would be ideal, to see if we can predict whether patients are likely respond to therapy or not. And again, leveraging the fact that we now have a way to predict what the phenotype of the tumor is. Of course, there's many different phenotypes. We've only looked at two right now in particular, but we know there are up to five different phenotypes of prostate cancers, so a lot more work to be done there.

Really, the clinical uses is eventually thinking about interventional trials, to stratify patients for specific therapies based on the predictive phenotype. We do need to get a handle on the additional ones too, because those also do exist in patients, although sometimes more rare. That's kind of where I see.

Andrea Miyahira: You mentioned in your introduction how complicated, painful, and sometimes just completely impossible it is for patients to get metastatic biopsies. How far do you think we are as a field from no longer needing metastatic biopsies and we can just use ctDNA or other liquid biopsy assays instead for making any patient treatment and management decisions?

Gavin Ha: Yeah, that's a tough question. I guess everything will be my opinion. I think ctDNA liquid biopsies really can offer additional insights for the oncologist and help with clinical care decisions. Certainly should be, if possible, done in parallel. If a biopsy is feasible, I don't think we would need to stop or replace that. Certainly you just take the ctDNA, see whether it adds additional value.

But, we still don't know whether ctDNA fully captures the extent of the heterogeneity. We're tackling that question. We have this proof of concept to look for phenotypic heterogeneity. But what can it offer in addition to or beyond what you would get from a single biopsy? If we're able to address that question more comprehensively, then I think we may have a better, clearer answer, or at least a path towards thinking about whether ctDNA may be able to replace the biopsy.

Of course, when a biopsy is not accessible, that I think that our work, and many others right now in the field, are really adding to the breadth of information that we can capture and extract from ctDNA. And I think that certainly goes a long way, and I think people who are caring for patients are starting to take notice to that. So, very happy to be part of that push and that drive.

Andrea Miyahira: All right. Well, thanks for joining us today and sharing this study. I look forward to seeing where you guys go next.

Gavin Ha: Thank you so much, Andrea.