The complexity and heterogeneity of cancer can be attributed to dynamic changes in tumour cell states1. This plasticity allows cancer cells to survive challenges such as chemotherapy and microenvironments encountered during metastasis2. Cancer cells maintain plasticity through aberrant activation of signalling pathways regulating epithelial (E) and mesenchymal (M) genes3–6. Defining cell positions within EM continua is key to understanding how cell states impact cancer progression and therapy escape7. Multiplexed imaging techniques provide detailed, multidimensional information about the molecular and cellular features of tumours. High-throughput imaging systems and automated data processing enable the extraction of precise molecular information, including cellular context and morphology, and protein levels, localisations, and interactions8. This molecular profiling may enable patient-specific cancer phenotyping and guide the application of cell state- targeted therapeutics9.
We used single-cell imaging to monitor the temporal evolution of lung adenocarcinoma cells exposed to cisplatin, doxorubicin, and TGF-B over an extended time course. Cells were fixed at 24, 48, 72 and 96h and individually labelled with markers of the epithelial-mesenchymal transition (EMT), including E-cadherin, N-cadherin, ZEB1, SNAI1, cytokeratin, CD44, and ZO-1, as well as proteins governing proliferation and the cell cycle, such as Ki67 and p21. Our findings reveal the temporal dynamics of therapy resistance, resulting in the suppression of proliferation via Ki67 silencing and the induction of senescence through p21 upregulation. We find that chemotherapy concurrently triggers an EMT-like response, and observed temporal differences in its onset between treatments.
For a more comprehensive analysis, we then applied our novel method of 'inferential multiplexing', involving deep image-to-image translation of label-free and fluorescence images to predict molecular marker labelling within cells. This allowed us to integrate information from all individually labelled EMT/cell cycle markers within each captured cell, resulting in a 9-plex single-cell image set. This data improved the classification of treatment-induced phenotypes beyond standard imaging, and pseudotime reconstruction further elucidated the kinetics of co-expressed proteins during phenotypic trajectories. Our study sheds light on the dynamic behaviour of cancer cells in response to therapy, highlighting the potential of artificial intelligence in advancing our understanding of cancer progression and treatment response.