Morphodynamic and motility feature-based deep learning classification for subtypes of cancer-associated fibroblasts.

Abstract

The cancer-associated fibroblasts (CAFs) have risen to prominence as key players in cancer progression. It is now widely appreciated that CAFs are composed of heterogeneous populations, either pro-tumorigenic or anti-tumorigenic. In addition, a growing body of evidence supports the multifaceted nature of CAFs. Therefore, researchers in the field have been trying to identify subtypes of CAFs with molecular markers. However, those markers cannot accurately classify the subpopulations and can be co-expressed in different subtypes. In order to utilize CAFs as a target for cancer treatment, issues with subtypes of CAFs must be resolved such that specific pro-tumorigenic subtypes can be suppressed or reprogrammed to anti-tumorigenic ones. The morphology and the motile characteristics of CAFs result from gene expression combinations. Thus, those characteristics can be holistic readouts of CAFs. Unlike biomolecular analysis, a fixed cell-based end-point assay, morphology or motility features of cells can be traced with live-cell imaging. Those features can provide information on the dynamic changes of CAFs. Here, in order to comprehensively identify subtypes CAFs, we adopt a deep learning-based cell classification strategy. Using multiple unsupervised and supervised machine learning algorithms, we extract the morphodynamic and motility features of cells from label-free live-cell imaging data of CAFs. To this end, we established in vitro breast CAFs, which were fibroblasts cocultured with two different breast cancer cell lines with different aggressiveness. As a result, we show that the morphodynamic and motility features can successfully classify heterogeneous subpopulations of cells.