Aim 1: Developing a panel to study the tumor-immune microenvironment in the draining LNs of melanoma patients
The panel included antibodies to identify tumor cells, fibroblasts, vasculature, different types of immune cells and functional markers implicated in immune evasion and response to immunotherapy. I have performed validation and calibration experiments to examine sensitivity and specificity of the antibodies. Each antibody was first tested by immunohistochemistry. Only antibodies of high sensitivity and specificity were chosen. I conjugated the validated antibodies to metals that allow MIBI-TOF detection, and then tested and titrated these conjugates in MIBI-TOF. Altogether, I devised a final panel of 39 antibodies that were used to stain our cohort.
Aim 2: Selecting samples for the cohort and data acquisition
We have collected a cohort of 80 sentinel LNs of melanoma patients at the time of diagnosis. Half of the patients had metastases in the sentinel LNs and half did not. In addition, they were retrospectively selected such that in each group half of them developed distant metastases and half did not within five years of diagnosis. We have collected the samples and created a tissue microarray (TMA) from the patients to reduce batch effects between patients and the quantity of antibodies. I have stained our data set with the validated antibody panel and acquired MIBI images of all our samples. Altogether 200 0.8x0.8mm2 images were acquired, encompassing ±2,000,000 cells.
Aim 3: Image analysis.
All images require a preprocessing step, intended to remove bleed-through from other channels, noise and antibody aggregates. We have developed Mass based imaging Analysis User Interface (MAUI), a graphical user interface that enables to perform these steps of imaging preprocessing (Baranski, Milo et al. Plos Comput Biol, 2021). Most of the images acquired in this project have passed already these steps and are ready for downstream analysis.
Aim 4: Cell classification
To perform downstream analysis in which we determine the organization of the tumor-immune microenvironment we need to identify individual cells in the images in a process named segmentation. To perform segmentation on our images we have used Mesmer, an algorithm that has automated this task (Greenwald et al., 2022). Following cell segmentation, the expression of each protein is quantified in each cell and is summarized in a cell table. This table serves as input for cell classification, where the type and phenotype of each cell in the tissue is inferred using the combination of co-expressed proteins, in combination with prior biological knowledge. For example, a cell that expresses CD3 and CD8 is classified as a CD8+ T cell. Cell classification methods typically use manual gating or clustering of the cell table.Inferring cell classifications from multiplexed images has unique challenges. Imaging artifacts are main challenges, and they include background noise and antibody aggregates. Biological factors also contribute to the difficulty of cell classification. In tissues and specifically in LNs cells are densely packed and cells are situated next to each other. In addition, cells extend cellular projection. These factors lead to signal spillover, whereby protein signals from one cell overlap with neighboring cells. All these issues make the task of cell classification challenging and time consuming and it requires sequential rounds of clustering, gating, and manual annotation by an expert. We have performed cell classification on 16 images from our dataset and identified twenty-four cell types including different types of immune cells, stromal cells, endothelial cells and tumor cells. The cells expressed their corresponding proteins. We have overlayed the different cell types on the images creating cell maps. Examination of these images revealed large diversity between patients .To improve the cell classification process, we have developed CellSighter (Amitay et al, bioRxiv 2022), a deep learning-based pipeline to perform cell classification in multiplexed images. CellSighter reaches 80-100% accuracy for most cell types. CellSighter can be applied across datasets thus leading to data integration and standardization. In summary CellSighter reduces significantly the time needed for cell classification and improves accuracy and consistency across datasets.
