ANALYSIS OF CELL MORPHOMETRIC PHENOTYPES BY DEEP LEARNING

Abstract

Cell morphology is profoundly influenced by cellular interactions with microenvironmental factors such as the extracellular matrix (ECM). Upon adhesion to specific ECM, various cell types are known to exhibit different but distinctive morphologies, suggesting that ECM-dependent cell morphological responses may harbour rich information on cellular signalling states. However, the inherent morphological complexity of cellular and subcellular structures has posed an ongoing challenge for automated quantitative analysis. Here we develop a deep learning-based analysis pipeline for the classification of cell morphometric phenotypes from multi-channel fluorescence micrographs, termed SE-RNN (residual neural network with squeeze-and-excite blocks). We demonstrate SERNN-based classification of distinct morphological signatures observed when fibroblasts or epithelial cells are presented with different ECM. We also demonstrate that SERNN-based feature vectors confer significant advantages in quantifying morphological differences over traditional geometric shape descriptors. Lastly, we also develop a generative framework to demonstrate the possibility of interpretation of deep learning-based feature vectors.