The spatial landscape of progression and immunoediting in primary melanoma at single cell resolution

Abstract

Cutaneous melanoma is a highly immunogenic disease, surgically curable at early stages, but life-threatening when metastatic. Here we integrate high-plex imaging, 3D high-resolution microscopy, and spatially-resolved micro-region transcriptomics to study immune evasion and immunoediting in primary melanoma. We find that recurrent cellular neighborhoods involving tumor, immune, and stromal cells change significantly along a progression axis from precursor states to melanoma in situ to invasive tumor. Hallmarks of immunosuppression are detectable by the precursor stage, and when tumors become locally invasive, a consolidated and spatially restricted suppressive environment forms along the tumor-stromal boundary. This environment is established by cytokine gradients that promote expression of MHC-II and IDO1 and by PDL1-expressing macrophages and dendritic cells engaging activated T cells. However, a few mm away, T cells synapse with melanoma cells in fields of tumor regression. Thus, invasion and immunoediting can co-exist within a few millimeters of each other in a single specimen.

UnMICST: Deep learning with real augmentation for robust segmentation of highly multiplexed images of human tissues

“…A total of ~10,400 nuclei were labelled by a human expert for nuclear contours, centers, and background. In addition, two human experts labelled a second dataset from a whole-slide image of human melanoma 34 to establish the level of inter-observer agreement and to provide a test data set that was disjoint from the training data.…”

“…A common approach judging the accuracy expected for a supervised learning method is to use multiple human experts to label the same set of data and determine the level of inter-observer agreement (of course, it may ultimately be possible to exceed this level of human performance 24, 29 ). We assessed inter-observer agreement using both the F1-score and sweeping IoU scores with data from whole-side images of human melanoma 34 . For a set of ~4,900 independently annotated nuclear boundaries, two experienced microscopists achieved a mean F1-score of 0.78 ( Supplementary Figure 1 ) and an IoU of 60% at a threshold of 0.6.…”

“…We assessed inter-observer agreement using both the F1-score and sweeping IoU scores with data from whole-side images of human melanoma 34 . For a set of ~4,900 independently annotated nuclear boundaries, two experienced microscopists achieved a mean F1-score of 0.78 (Supplementary Figure 1) and an IoU of 60% at a threshold of 0.6.…”

Visinity: Visual Spatial Neighborhood Analysis for Multiplexed Tissue Imaging Data

“…Other approaches compute groups (topics or clusters) based on probabilities and distances. Zhu et al [83] use a Hidden-Markov random field to model spatial dependency of gene expression, and Spatial-LDA [75] is a probabilistic topic modeling approach, which is also applied in the biomedical field [48]. Most similar to our approach, stLearn [53] and CytoMAP [71] rely on traditional distance-based clustering.…”

In situ functional cell phenotyping reveals microdomain networks in colorectal cancer recurrence

“…Single-cell approaches, which include a “large” set of biomarkers (including DNA, RNA, and proteins), are well suited to capture this emergent phenotypic continuum ( Nachmanson et al, 2021 ; Nirmal et al, 2021 ). However, preserving the spatial context is equally crucial to capture the various functional states that cells might emerge through their neighborhood interactions ( Vitale et al, 2021 ; Zanotelli et al, 2020 ).…”