Aim 1: Use MIBI-TOF to chart the iTME in clinical samples from melanoma patients and delineate its function in response to different immunotherapies.
In collaboration with Scott Rodig from the Dana Farber Cancer Institute, we compiled a cohort from 80 patients with stage III and IV metastatic melanoma, including up to five regions per patient (>300 samples). We have collected metastases from distinct anatomical sites, including lymph nodes, lungs, skin, liver and brain and profiled them using multiplexed imaging. We then used the specialized computational tools that we developed to analyze this data, including low-level analysis, cell segmentation, cell classification and microenvironmental analysis. We identified a hierarchy of immune infiltration into the tumor, whereby immune populations cooccur together across patients. This hierarchy mirrors previous results that we have obtained in Breast Cancer , suggesting that it may be a general principle organizing the immune microenvironment. We also identified systematic changes between anatomical sites, for example lung metastases overall have increased immune infiltration. Work is ongoing to identify predictive features of response to immunotherapy, as well as specific immune populations that characterize specific anatomical sites.
Aim 2: Profile murine melanoma tumors to elucidate genetic and temporal mechanisms that drive iTME organization in vivo
In collaboration with the lab of Yardena Samuels, we are studying how tumor heterogeneity affects the tumor iTME in melanoma. We implanted mice with tumors composed of either single-cell clones (SCC), mixture of clones (12mix) or the original parental cell line (UVB), and harvested tumors and lymph nodes at different time points. Altogether we collected ~200 tumors and lymph nodes for spatial and temporal analysis.
We found that homogenous tumors were quickly rejected whereas heterogeneous tumors grew, and tumors of the heterogeneous parental cell line (UVB) were significantly more aggressive. Comparing UVB to SCC tumors, we demonstrate that rejected homogenous tumors are highly infiltrated with immune cells, whereas growing, heterogeneous tumors are significantly less infiltrated. Rejected SCC tumors demonstrate F4/80+ macrophages infiltration as early as 6 days following implantation. By day 10, these tumors exhibit massive infiltration of T cells and CD11b+/F4/80+ macrophages prior to tumor rejection. However, growing 12mix and UVB tumors, exhibit accumulation of T cells by day 13, followed by a significant drop in the number of infiltrated T cells between days 13 and 27. Interestingly, in the highly aggressive UVB tumors, this is accompanied by loss of CD11b expression in F4/80 macrophages. These findings demonstrate coordinated temporal immune response that depends on clonal heterogeneity. Additional analysis will reveal the specific immune-cell subtypes involved in this response.
Aim 3: Develop new experimental tools for multiplexed imaging
There is a critical unmet need to develop tools that will allow to mark and trace single cells in tissues while surveying a high degree of information from single cells. In collaboration with the group of Lucia Gemma Delogu, we developed a versatile multiplexed label‐free single‐cell detection strategy based on single‐cell mass cytometry (CyTOF) and MIBI‐TOF. We use 2-D materials, MXenes, to label cells in vitro. We then use MIBI-TOF to detect MXenes in different organs revealing their spatial distribution. We are currently expanding on this approach to follow infused T cells infiltrating into tumors.
Aim 4: Develop new machine-learning-based algorithms to analyze this novel data
We have developed CellSighter, a deep-learning based pipeline to perform cell classification in multiplexed images (PMID 37463931). CellSighter bypasses the challenges of tabular data and uses state-of-the-art computer vision deep learning algorithms to work directly on the images, which contain information that is critical for classification. Given a small training set of expert-labeled images, CellSighter outputs the label probabilities for all cells in new images. We tested CellSighter on data from different multiplexed imaging modalities, including MIBI-TOF, IMC and CODEX, and found that it achieves 80-100% accuracy for major cell types, which approaches inter-observer concordance.
