The transition of epithelial cells into a mesenchymal state (epithelial-to-mesenchymal transition or EMT) is a highly dynamic process implicated in various biological processes. During EMT, cells do not necessarily exist in ‘pure’ epithelial or mesenchymal states. There are cells with mixed (or hybrid) features of the two, which are termed as the intermediate cell states (ICSs). While the exact functions of ICS remain elusive, together with EMT it appears to play important roles in embryogenesis, tissue development, and pathological processes such as cancer metastasis. Recent single cell experiments and advanced mathematical modeling have improved our capability in identifying ICS and provided a better understanding of ICS in development and disease. Here, we review the recent findings related to the ICS in/or EMT and highlight the challenges in the identification and functional characterization of ICS.

Rapid growth of single-cell transcriptomic data provides unprecedented opportunities for close scrutinizing of dynamical cellular processes. Through investigating epithelial-to-mesenchymal transition (EMT), we develop an integrative tool that combines unsupervised learning of single-cell transcriptomic data and multiscale mathematical modeling to analyze transitions during cell fate decision. Our approach allows identification of individual cells making transition between all cell states, and inference of genes that drive transitions. Multiscale extractions of single-cell scale outputs naturally reveal intermediate cell states (ICS) and ICS-regulated transition trajectories, producing emergent population-scale models to be explored for design principles. Testing on the newly designed single-cell gene regulatory network model and applying to twelve published single-cell EMT datasets in cancer and embryogenesis, we uncover the roles of ICS on adaptation, noise attenuation, and transition efficiency in EMT, and reveal their trade-off relations. Overall, our unsupervised learning method is applicable to general single-cell transcriptomic datasets, and our integrative approach at single-cell resolution may be adopted for other cell fate transition systems beyond EMT.

Epithelial-to-mesenchymal transition (EMT) plays an important role in many biological processes during development and cancer. The advent of single-cell transcriptome sequencing techniques allows the dissection of dynamical details underlying EMT with unprecedented resolution. Despite several single-cell data analysis on EMT, how cell communicates and regulates dynamics along the EMT trajectory remains elusive. Using single-cell transcriptomic datasets, here we infer the cell–cell communications and the multilayer gene–gene regulation networks to analyze and visualize the complex cellular crosstalk and the underlying gene regulatory dynamics along EMT. Combining with trajectory analysis, our approach reveals the existence of multiple intermediate cell states (ICSs) with hybrid epithelial and mesenchymal features. Analyses on the time-series datasets from cancer cell lines with different inducing factors show that the induced EMTs are context-specific: the EMT induced by transforming growth factor B1 (TGFB1) is synchronous, whereas the EMTs induced by epidermal growth factor and tumor necrosis factor are asynchronous, and the responses of TGF-β pathway in terms of gene expression regulations are heterogeneous under different treatments or among various cell states. Meanwhile, network topology analysis suggests that the ICSs during EMT serve as the signaling in cellular communication under different conditions. Interestingly, our analysis of a mouse skin squamous cell carcinoma dataset also suggests regardless of the significant discrepancy in concrete genes between in vitro and in vivo EMT systems, the ICSs play dominant role in the TGF-β signaling crosstalk. Overall, our approach reveals the multiscale mechanisms coupling cell–cell communications and gene–gene regulations responsible for complex cell-state transitions.

How basal cell carcinoma (BCC) interacts with its tumor microenvironment to promote growth is unclear. We use singe-cell RNA sequencing to define the human BCC ecosystem and discriminate between normal and malignant epithelial cells. We identify spatial biomarkers of tumors and their surrounding stroma that reinforce the heterogeneity of each tissue type. Combining pseudotime, RNA velocity–PAGA, cellular entropy, and regulon analysis in stromal cells reveals a cancer-specific rewiring of fibroblasts, where STAT1, TGF-β, and inflammatory signals induce a noncanonical WNT5A program that maintains the stromal inflammatory state. Cell-cell communication modeling suggests that tumors respond to the sudden burst of fibroblast-specific inflammatory signaling pathways by producing heat shock proteins, whose expression we validated in situ. Last, dose-dependent treatment with an HSP70 inhibitor suppresses in vitro vismodegib-resistant BCC cell growth, Hedgehog signaling, and in vivo tumor growth in a BCC mouse model, validating HSP70’s essential role in tumor growth and reinforcing the critical nature of tumor microenvironment cross-talk in BCC progression.

Lysine crotonylation (Kcrot), as a post-translational modification (PTM) originally identified at histone proteins, is involved in diverse biological processes. Several conventional machine-learning (ML) predictors were developed based on the Kcrot sites from histone proteins. Recently, thousands of Kcrot sites have been experimentally verified on non-histone proteins from multiple species. Accordingly, a few predictors have been developed for predicting the Krot sites for specific organisms (i.e. humans and papaya). Nevertheless, there is a lack of research on the comparison of the crotonylomes of different organisms. Here, we collected around 20,000 Kcrot sites experimentally identified from four different species as the benchmark data set. We present the deep-learning (DL) architecture dubbed DeepKcrot for predicting Kcrot sites on the proteomes across various species. DeepKcrot includes species-specific and general classifiers using a convolutional neural network with the word embedding (CNNWE) encoding approach. CNNWE performs better than both the traditional ML-based and other DL-based classifiers in terms of ten-fold cross-validation and independent test, independent of the size of the training set. Additionally, cross-species performance for each species-specific predictor is not as good as the self-species performance whereas the cross-species performance generally increases with the size of the training dataset. Moreover, the predictors developed based on the non-histone Kcrot sites are unsuccessful for the histone Kcrot prediction, suggesting that the Kcrot-containing peptides from non-histone and histone proteins have significantly different characteristics and data integration is required. Overall, DeepKcrot is an efficient prediction tool and freely available at http://www.bioinfogo.org/deepkcrot.