By Deborah Borfitz
May 6, 2020 | Researchers in the lab of Sangeeta Bhatia at the Massachusetts Institute of Technology (MIT) have spent the past several years developing nanoparticle sensors that can detect cancer by interacting with enzymes called proteases that help tumor cells escape from their original locations. Most recently, they have demonstrated the potential of coupling the approach with machine learning to develop a noninvasive, radiation-free and relatively inexpensive urine test as a complement to CT lung cancer screening.
If results in mice are indicative of what is possible in humans, a dipstick may one day be a viable alternative to standard CT screens in resource-poor settings around the globe, according to MIT and Harvard University graduate students Jesse Kirkpatrick and Ava Soleimany, and former MIT graduate student Andrew Warren, lead authors of the newly published study (DOI: 10.1126/scitranslmed.aaw0262) appearing in Science Translational Medicine. Their new test, powered by a panel of nanoparticle sensors targeting proteases dysregulated in lung cancer, proved more accurate than high-resolution CT scans in detecting localized lung cancer in mice as well as distinguishing cancer from benign inflammation in the lungs.
“Machine learning is really the best way of integrating multiple signals together to provide higher diagnostic accuracy compared to any one signal,” which has “emerging appreciation” in the field, says Kirkpatrick. He points to the multi-marker approach of the popular Cologuard test, which combines DNA mutation and protein biomarkers in the stool to screen for colon cancer. Other companies and research groups have had success uniting circulating DNA and protein biomarkers in the blood to detect cancer.
Unlike current diagnostic tests that measure endogenous biomarkers, the one under development in Bhatia’s lab can provide a functional measurement of the disease microenvironment, says Kirkpatrick, which might offer a better readout on disease activity. That would be particularly useful when it comes to distinguishing between bacterial and viral pneumonia—another area of investigation for the nanoparticle sensors.
“A large portion of our lab is currently devoted to advancing the sensors for lung cancer detection toward clinical implementation,” Kirkpatrick says. While one team is developing clinically relevant delivery methods for the sensors, others are working on rapid, inexpensive ways to detect the urinary reporters at the point-of-care. “We’re also interested in leveraging nanoparticles for other monitoring strategies, including assessing whether patients are responding to their treatment.”
Clever Workaround
Proteases are a class of enzymes that can cut and degrade other proteins. These enzymes play key roles in cancer progression, including the recruitment of blood vessels and metastasis, and are what allow cancer cells to spread and invade surrounding tissue by cutting proteins of the extracellular matrix. “But these proteases don’t really escape from the tumor site themselves,” Kirkpatrick says.
In the tumor microenvironment, locally acting proteases can cut off short protein fragments (aka peptides) present on the surface of the nanoparticles, Kirkpatrick explains, releasing synthetic reporter molecules small enough to escape through the bloodstream and ultimately be cleared by the kidneys into the urine.
The principal issue with low-dose CT scans—the gold standard for lung cancer screening—is that 96% of what gets detected prove to be false positives, says Kirkpatrick. “This is a problem because many of these patients have to undergo invasive follow-up procedures like biopsy or even tumor resection, and this comes with a risk of pneumothorax, or lung collapse.” In many cases, patients die from those confirmatory diagnostic tests.
The supporting evidence includes results of the National Lung Screening Trial, published in the New England Journal of Medicine, which investigated whether CT screening for lung cancer in high-risk subjects could reduce mortality from lung cancer. A follow-up study focused on the cost-effectiveness of screening found the cost to be “quite high” at $81,000 per quality-adjusted life-years gained, Kirkpatrick notes.
Mice To Men
In selecting the proteases to include in the new test, the MIT researchers began by looking at existing datasets from analyses of biopsied tumors from human lung cancer patients (The Cancer Genome Atlas) as well as from mouse models of lung cancer, says Soleimany. “From that we were able to nominate a panel of proteases that are dysregulated in lung cancer, and then engineered a panel of 14 nanoparticles that could be efficiently recognized and cleaved by those proteases.”
Each of the 14 nanoparticles was validated in the lab by designing sequences of the peptides that get cut by their target protease and then performing an assay to measure how well those sequences were recognized and cleaved by the nominated proteases, Soleimany says. “Moving forward, we’re interested in looking at how to further [enhance] the design of these nanoparticles for improving the sensitivity and specificity of our diagnostic test.” One possibility is by examining panels of hundreds or thousands of short recognition sequences against panels of protease candidates to identify the best ones to add to the nanoparticle sensors.
The mode of delivery will also need to change before the test could move into clinical trials, Kirkpatrick adds. In the experiments with mice, researchers inserted a small tube in the tip of the trachea and administered the nanoparticles to the animals’ lungs. A more clinically relevant delivery method for humans would be through a dry powder inhaler or nebulizer.
“This was the first time we delivered these nanoparticles directly into the lung via intratracheal administration, and the reason was that we wanted to concentrate our nanoparticles directly at the tumor site,” he says. “And it was very effective—almost 100% of the nanoparticles we delivered into the lungs remained [there] whereas with intravenous administration only about 20% of nanoparticles end up in the lungs, and the rest are delivered to other organs like the liver and kidneys.”
This was a key advance, Kirkpatrick continues, allowing researchers to detect tumors as small as about 3 cubic millimeters.
Laudable Performance
The initial test of the new nanoparticle sensors involved mice genetically engineered to develop lung cancer, allowing researchers to monitor the size of tumors over time in a more human-relevant manner, says Kirkpatrick. The ability of the sensors to measure the tumors happened at three time points, in parallel with high-resolution CT scans capable of detecting extremely small tumors.
As early as the 7.5-week time point, tumors “as small as the head of a pin” could be reliably detected in mice using the sensors, he says. At that point, tumors are on average only 2.8 cubic millimeters in size.
Whether or not this translates into early stage lung cancer detection in humans remains to be proven, Kirkpatrick adds. “But the tumors were extremely small and hadn’t metastasized to other organs and at the 7.5-week time point our sensors were more accurate than the corresponding CT scans.”
They could also reliably distinguish lung cancer from noncancerous inflammation in the lungs—another potential advantage over CT scans, says Kirkpatrick. And that can be attributed to the fact that the sensors incorporate 14 nanoparticles that can collectively detect a wide range of proteases implicated in lung cancer.
Disease Signatures
Machine learning was used to develop classifiers that could differentiate disease states based on the urinary readout generated by the nanoparticle sensors. “We saw this as a proof of concept that we could distinguish between lung cancer and benign lung disease, but clinically benign tumors might represent anything from a granuloma to a slightly enlarged lymph node,” Kirkpatrick says.
As Soleimany explains, this was a first-time use of machine learning for this purpose. The analysis work involved writing a computer algorithm to recognize patterns in the urinary data produced by the 14 sensors that corresponded to healthy mice or benign inflammation and to those with lung cancer. The resulting algorithm could then be applied to the urine data from subsequent animals to generate a prediction.
Machine learning will now be applied to other development efforts in their lab to generate more disease signatures, she adds. Researchers in Bhatia’s lab have previously developed nanoparticle sensors to detect a variety of conditions, including colon and ovarian cancer as well as coagulation disorders and infectious diseases.
The lab is currently working on sensors that can distinguish between bacterial and viral pneumonia, which could help doctors to determine which patients need antibiotics, Soleimany says. If successful, they could provide complementary information to nucleic acid tests like those being developed for COVID-19 by directly measuring the activity of the pathogen in the lungs and helping to distinguish active infection from disease exposure.
Glympse Bio, a company co-founded by Bhatia, is also working on developing this approach to replace biopsy in the assessment of liver disease. Neither Soleimany nor Kirkpatrick have any involvement in the company.