空间生物学工具识别肺腺癌中可预测生存的空间定位不同的成纤维细胞

Predicting overall patient survival from excised lung cancer tissue is a challenging task. We applied spatial biology tools to define 10 cell types using unsupervised clustering and using the spatial distributions of these 10 cell types and unsupervised clustering to define 8 cell neighborhoods to over 400 Adenocarcinoma TMA spots. The unsupervised cell type clustering identified 4 types of tumour/epithelial cells, 4 types of immune like cells and 2 types of potentially CAF cells. The frequencies of the cell types within the TMA spots correlated with a host of clinical variables such as stage, tumor size, differentiation degree, EGFR mutation status, patient sex, etc. As did the frequencies of the cells in the 8 neighborhoods. Collapsing the 8 neighborhood types into 3 neighborhoods (tumor, stroma and cells boarding tumor and stroma) and selecting the 2 types of CAF cells identified we calculated the frequency and density of CAF cells in the three neighborhoods. We found that the density and frequency of CAFs in the stroma neighborhood was highly predictive of overall survival in early stage (< 1B) Non smokers (p=0.0005 females, p=0.00007 males) but not as strong in early stage smokers. For Late stage lung cancers (>=1b) the frequency of a type of large tumor cell combined with the density (number of stromal cells per mm2) was highly predictive (p=0.000005) of outcomes for late stage current smokers in both males and females. Key to the success of this analysis was the exact segmentation of the all the DNA specific stained nuclei, even in areas of highly overlapping nuclei, using a novel deep learning enabled segmentation that allow pixels to belong to more than one nucleus.