![The left column shows gold plasmonic nanoparticles bound to the ovarian cancer marker CA125, while the remaining columns show the multiple fluorescent labels serving to identify the type of immune cell [2].](https://img.laserfocusworld.com/files/base/ebm/lfw/image/2021/09/2109LFW_wn_1.6139286b9803e.png?auto=format,compress&fit=max&q=45&w=250&width=250 "The left column shows gold plasmonic nanoparticles bound to the ovarian cancer marker CA125, while the remaining columns show the multiple fluorescent labels serving to identify the type of immune cell [2].")
New capabilities arise when incremental improvements across a variety of technologies combine to enable new applications. Recent work by a team of researchers at the University of Wisconsin–Madison (UW; Madison, WI), PNP Research Corporation (Drury, MA), and other academic and industrial partners embodies that pattern. They integrated developments in artificial intelligence (AI), automated image acquisition, plasmonic signal enhancement, and other technologies to create a potential ovarian cancer detection method.
According to the American Cancer Society, only about 20% of ovarian cancers can be diagnosed early.1 Although ovarian cancers overexpress a specific protein, MUC-16, serum MUC-16 levels don’t correlate with the disease. This new research shows a correlation between immune system cell binding patterns of MUC-16 and the presence or absence of ovarian cancer.2
A three-pronged problem
MUC-16 is a transmembrane glycoprotein belonging to the mucin family—molecules thought to protect against infectious agent transmission at mucosal surfaces. Enzymes sever the molecule from its membrane connection, so it’s reasonable to expect higher levels of expression would lead to higher serum concentrations. MUC-16 has a repeating epitope labeled CA125; detection methods generally target this epitope, being expressed, for example, as CA125 serum levels, which don’t correlate well with disease progression.
In previous work, the UW-led team had shown that CA125 bound at high levels to peripheral blood mononuclear cells (PBMCs)—a variety of cell types essential to the human immune system. The next step was to evaluate the binding levels for each of the different PBMC types. They faced three problems: a single PBMC will bind only on the order of a few MUC-16 molecules; meaningful statistics require cytometry on large numbers of cells; and different cell populations need to be reliably identified.
To reach the required low detection limit, they incubated the sample with 80 nm gold nanoparticles conjugated to anti-CA125 antibodies (see figure). Under dark-field illumination, the gold nanoparticles are 13X brighter than the intracellular background illumination. In earlier work, the team had shown that an average cell-bound MUC-16 strand would most likely bind three or fewer gold nanoparticles.3 Clusters red-shift, so they can be spectrally distinguished from single nanoparticles. The task reduces to counting the number of gold nanoparticles on each cell. In practice, there are multiple challenges, starting with identifying the boundaries of cells in a population of varying amorphous shapes.
For a screening assay, it’s impractical for a human observer to identify individual cell boundaries, so the researchers turned to AI. Lab personnel identified boundaries for 3192 PBMCs and used that set to train, validate, and test an area-matching model. Acquired images were processed by the model, then screened to eliminate cells smaller than 16 µm2 or larger than 400 µm2, which are either artifacts or cells not of interest. The third problem, sorting the different PBMCs by cell type, was solved with multiple fluorescent labeling—a solution intimately tied to the design of the optical system itself.
Multiple optical functions for automated screening
PBMCs come in several types, including T-cells, B-cells, NK-T cells, NK cells, and monocytes. Each of them express a unique combination of cluster-of-differentiation (CD) protein markers. The researchers incubated the samples with different fluorescent labels for CD45 (a marker present in all leukocytes), CD19, CD56, CD3, and CD14. By sequentially rotating through four-channel excitation and emission filters, they identified the five labels and classified the cells. That was in conjunction with an automated image acquisition process that translates the sample to acquire 105 different fields of view, automatically identifies and focuses at the optimum imaging plane in each of the four color bands, switches to dark-field illumination, and then translates in 0.5 µm z increments to create a 40-slice image stack. Acquisition of the full 105-field 3D image stack takes 58 minutes. Identification of the fluorophores (and, therefore, the cell type), delineation of cell borders, and nanoparticle counting are automated, creating a suitable screening process.
In this proof-of-principle exploration, the team evaluated serum samples from 14 serous ovarian cancer patients and seven healthy controls. The average number of nanoparticles bound to B-cells was more than 10 for patients and around five in healthy controls, while T-cells in patients bound more than four nanoparticles, compared to about one in controls. Cell binding was not correlated with serum CA125 levels, showing that this level of advanced automated cytometric analysis provides a degree of clinical insight not otherwise available.
“This fully automated microscopy system,” the researchers stated, “now enables new studies to examine the relevance of such cell-surface molecules in the course of disease.” Not only for ovarian cancer, as demonstrated in this work, but it’s “readily adapt[able] to the quantification of other cell surface molecules that are relevant in the course of disease.”
REFERENCES
1. See https://bit.ly/Gaughan-Ref1.
2. G. González et al., Cancers, 13, 2072 (2021); doi:10.3390/cancers13092072.
3. S. Jeong et al., ACS Sens., 5, 9, 2772–2782 (2020); doi:10.1021/acssensors.0c00567.
