Credit: Lehigh University
Every year, 14,000 women in the United States die from ovarian cancer. It’s the fifth most common type of cancer that kills women, and part of the reason it’s so deadly is that it’s hard to find in its early stages. Patients usually don’t notice any signs of cancer until it has spread, and there are no reliable screening tests to find it early.
A group of scientists is trying to change that. Researchers from the Memorial Sloan Kettering Cancer Center, Weill Cornell Medicine, the University of Maryland, the National Institutes of Standards and Technology, and Lehigh University are part of the group.
Two papers that came out recently talk about their progress toward a new way to find ovarian cancer. Machine learning techniques are used to quickly analyse the spectral signatures of carbon nanotubes to find biomarkers of disease and to identify the cancer itself.
In November, the first paper came out in Science Advances.
Yoona Yang, a postdoctoral research associate in Lehigh’s Department of Chemical and Biomolecular Engineering and co-first author of the paper with Zvi Yaari, a postdoctoral research fellow at Memorial Sloan Kettering Cancer Center in New York, says, “We showed that a perception-based nanosensor platform could detect ovarian cancer biomarkers using machine learning.” Ming Zheng, a research chemist at the National Institute of Standards and Technology, Anand Jagota, a professor of bioengineering and chemical and biomolecular engineering at Lehigh University, and Daniel Heller, an associate member at the Memorial Sloan Kettering Cancer Center and head of the Cancer Nanotechnology Laboratory, were also authors.
Jagota, who is also the associate dean of research for Lehigh’s College of Health, and Yang are both part of Lehigh’s Nano | Human Interfaces Presidential Initiative. This is a multidisciplinary research project that aims to change the way we work with data and the complex tools of scientific discovery.
In the past, a molecular recognition molecule like an antibody had to be matched to each biomarker in order to find disease. But there isn’t a single biomarker or analyte that shows that someone has ovarian cancer. When more than one analyte needs to be measured in a given sample, which can make a test more accurate, more antibodies are needed. This makes the test more expensive and takes longer to finish.
Yang says, “Perception-based sensing works like the brain.” “The system is made up of a sensing array that picks up a certain feature of the analytes in a certain way. The computational perceptive model then looks at the ensemble response from the array. It can find multiple analytes at the same time, which makes it much more useful.”
In this study, the array was made of single-wall carbon nanotubes that were wrapped in strands of DNA. Because of how the DNA was wrapped and the different DNA sequences that were used, the surfaces of the nanotubes were all different. In turn, the different surfaces attracted a variety of proteins from a uterine lavage sample that was rich in different levels of biomarkers for ovarian cancer.
Heller says that the electronic properties of carbon nanotubes are interesting. “When you shoot light at them, they give off a different colour of light, and the colour and strength of that light can change depending on what’s stuck to the nanotube. By using a variety of nanotubes with different wrappings, we were able to take advantage of the complexity of so many possible binding interactions. And that gave us a variety of sensors that could all detect slightly different things, and it turned out that they reacted differently to different proteins.”
The spectral signatures were used to teach the machine learning algorithm how to recognise the pattern of nanotube emission that showed the presence and amount of each biomarker.
“The big idea here is that these nanotubes can be used as sensors for anything,” says Jagota. “They don’t know anything about biomarkers, which means they’re not set up to bind to anything in particular. We only knew that they could be put in an aqueous medium and that whatever was in that medium would cause spectral shifts and changes in intensity. And by using a mix of these sensors, we were able to train the algorithm to convert these inputs to outputs in a very accurate way using math. It’s like having 20 pairs of eyes that see different things at the same time. No single eye is that good, but a group of them can be trained to be better at finding ovarian cancer than what we have now.”
Many of the same researchers worked on the second paper, which came out in March in Nature Biomedical Engineering. YuHuang Wang, a professor in the University of Maryland’s department of chemistry and biochemistry, and Mijin Kim, a postdoctoral research associate at Memorial Sloan Kettering Cancer Center and the study’s lead author, were also authors.
Heller says, “In this paper, we didn’t look at biomarkers; instead, we looked at the disease itself.” “We wanted to know if this technology could tell the difference between a blood sample from a person with ovarian cancer and one from a person without the disease.”
There were both healthy people and people with other diseases among the patients who did not have ovarian cancer.
In this study, quantum defects were added to the nanotubes to make them work better. This made the nanotubes respond in a wider range of ways.
“A certain molecule was attached to the nanotubes, which gave them an extra data signal,” says Jagota. “So, every nanotube-DNA combination led to more information. The disease state, not the biomarker, was used to train the model.”
The spectral emissions of the nanotubes were used to create a “fingerprint” of the disease. The results were statistically significant when it came to how well the model could find ovarian cancer and how well it could find both known and unknown biomarkers of the disease.
In both papers, Heller uses the nose as a way to explain how the machine learning model works. For example, not every smell has its own odour receptor.
“Instead, there are many different smell receptors that bind to certain molecules and make a pattern, or something like a fingerprint,” he says. “And your brain analyses that pattern, which tells you what you’re smelling. So, there isn’t a single sensor that only responds to a single thing. But the algorithm can tell what’s a biomarker and what’s not based on how the different sensors respond with different changes in colour and wavelength intensity.”
The team has shown that their method can find ovarian cancer better than other methods, but it still can’t find the disease in its early stages. Heller says that part of the problem is that so few people are diagnosed at those times that it is hard to find enough samples to train the algorithm.
“We’re trying to figure out how to find this disease as early as possible,” he says.
Jagota says that the next steps could include making the method work for more diseases and figuring out if it can be improved to work in a clinical setting.
“And this is a method that can be used in a lot of different situations,” he says. “We’re mostly concerned with health, but it could also be used to find pollution in the air, for example. I find it interesting that it could be used to treat a wide range of diseases and conditions.”
Further information: Zvi Yaari et al, A perception-based nanosensor platform to detect cancer biomarkers, Science Advances (2021). DOI: 10.1126/sciadv.abj0852
Mijin Kim et al, Detection of ovarian cancer via the spectral fingerprinting of quantum-defect-modified carbon nanotubes in serum by machine learning, Nature Biomedical Engineering (2022). DOI: 10.1038/s41551-022-00860-y
Journal information: Nature Biomedical Engineering , Science Advances
Source: Lehigh University