Conversations With Prostate Cancer Experts

Leave a comment

Diagnosing Cancer Via Liquid Biopsy and Machine Learning

Ms. Jina Ko is a PhD student in the Department of Bioengineering at the University of Pennsylvania. She was among 14 PhD candidates from the U.S., Canada, and Germany to be named to the inaugural class of Schmidt Science Fellows. Ms. Ko works in the lab of Professor David Issadore on microfluidics and lab-on-a-chip technologies.

Dr. David Issadore is an Assistant Professor of Bioengineering and Electrical and Systems Engineering at the University of Pennsylvania. Dr. Issadore’s research focus is on applying microelectronics, microfluidics, nanomaterials, and molecular targeting to medicine. His lab explores how these new technologies can bring medical diagnostics from expensive, centralized facilities directly to clinical and resource-limited settings.

Ms. Ko and Dr. Issadore spoke with Prostatepedia about a platform for diagnosing pancreatic cancer via liquid biopsy and machine learning, a technology that can be applied to other cancer types, including potentially prostate cancer.

Not a member? Join us!

What drew each of you to the world of bioengineering?

Ms. Jina Ko: I did a lot of internships as an undergraduate student working on biomechanics, point-of-care diagnostics, and all things cancer biology. I got interested in diagnostics because, even though there are goodcurrent treatment options and emerging treatments, if we don’t have good diagnostics to guide patients to the right treatment options, they cannot really benefit. Good diagnostics that guide patients can be a huge bridge to connect patients to treatments.

In terms of pancreatic cancer, everyone’s diagnosed really late, when they already have metastases.

I saw that as a good chance to develop early stage pancreatic cancer diagnostics: to detect them before metastasis, so that we can increase the survival rate.

Dr. David Issadore: My training is in physics and electrical engineering. I trained to design computer chips and got into diagnostics because I became interested in whether or not the same approaches that reduced costs could be applied to medicine.

In the 1960s, electronics were only accessible to big institutions and people with a lot of money, but now, everyone has access to cellphones and laptops. I was interested in whether or not we could do the same thing for medicine, to make ultrasensitive diagnostics that do nearly impossible things and solve intractable problems by miniaturizing and integrating them. That’s what we do in my lab here at the University of Pennsylvania.

For her PhD, Jina had a brilliant insight into a new device. She took it all the way from a drawing on the back of an envelope to something we use on patient samples to diagnose disease. She did that in five years, which is rare and pretty incredible.

You’ve worked on two main projects: integrating microchip-based technologies with machine learning for liquid biopsies and integrating nanofluidic technology with machine learning to diagnose cancer. Can you tell us a bit about that work?

Ms. Ko: Our platform is a combination of those two projects: our approach looked at liquid biopsies to find blood-based biomarkers so that we can minimize invasion for biomarkers rather than doing invasive biopsy. For biomarkers, we focused on exosomes, which are small particles that circulate in the bloodstream. Exosomes are great as biomarkers, because they have good molecular information of their mother cells. For example, it’s really hard to get at pancreatic cancer cells because of invasive biopsy. But we can derive pancreatic cancer cell exosomes from the blood. The challenge is that they’re really small, on the nanoscale at only 100 nanometers in diameter.

So, we need a good tool to isolate those exosomes and profile them for the molecular signature.

At first, we developed tools that can isolate specific types of exosomes, so that we can enrich the exosomes

and profile them. Even though we profiled the exosomes, we noticed that if we just look at one expression level, with molecular cargo like DNA or RNA inside, we can cover the heterogeneity of different patients.

Also, pancreatic cancer is heterogeneous. That’s why we used machine learning rather than profiling individual RNAs from exosomes. We thought we could find a pattern, a combination of biomarkers so that we can find orthogonal information and the signature inside. We applied machine learning to decrease multidimensions into a single score.

That score can then tell us whether a person has pancreatic cancer.

And how accurate was it?

Ms. Ko: We started with three mouse model groups. One group was healthy, one had tumors, and the third group had lesions in the pancreas, but they did not yet have a tumor, which is considered pre-cancerous.

In the three-way comparison, we got 100% accuracy, but it’s a small size sample. There were only about 20 mice, so we definitely need to increase the number to ensure that it’s an accurate representation. We applied it to clinical patients where we classified metastatic pancreatic cancer patients to healthy controls. In that study of 24 patients, we got 100% accuracy as well.


Ms. Ko: Yes. Even though it’s a small sample size, we got extremely accurate molecular signatures from exosomes. We really want to apply this to earlystage pancreatic cancer patients, to screen some risk groups. We want to be able to predict if people were going to develop pancreatic cancer at a later stage before the disease appears.

Dr. Issadore: The next step of early detection in humans is challenging because we need to measure a lot of people and only some of them are going to get cancer. This study lays the groundwork to take that next step, and we’re gearing up to do that.

Do you have any plans to study different types of cancer, or will you just continue looking at pancreatic cancer?

Dr. Issadore: No. We want to branch out. Every cell in the body sheds these exosomes, and the machine learning approach allows us to look for signatures without having to understand the underlying biology. This means that, as long as there is a signature—a difference between cells that are cancerous and cells that are healthy—this technique should work for any type of cancer.

It will be a challenge to find the right animal and clinical models to develop early detection, so we’re working with collaborators at the Abramson Cancer Center. Together, we will link this technology with models of breast cancer, leukemia, and many others.

We’ve also taken the same approach and applied it to different diseases. We’ve tried it with traumatic brain injury and have had exciting results. It’s a pretty general technique.

Ms. Ko: Whenever we talk about 100% accuracy in science, people are a little suspicious because it’s pretty rare. To validate that level of accuracy from the pancreatic cancer molecular signatures that we found, rather than some random artifact from machine learning, we trained the algorithm with wrong labels. We shuffled the labels and eliminated the molecular signatures on purpose, and then we trained the algorithm to make sure that it failed.

One hundred percent does sound too good to be true.

Dr. Issadore: You have to do a lot of controls, which we did. It wouldn’t be 100% if we had 1,000 or 10,000 samples. But for the members we tested, it was perfect.

You said the obstacles to moving forward are just getting enough people to test? Any other obstacles to the process?

Ms. Ko: Increasing the sample size can be one option, but we want to also find a subgroup classification that can help with clinical decisions. We are looking at short survival versus long survival patients to find signatures there. We are also looking at metastatic patients versus no visible metastases to better understand metastases.

Not a member? Join us to read more about tech + health collaborations for cancer.