Objective/Rationale
Fluorescence microscopy has permitted interrogation of molecular mechanisms, high-throughput phenotypic drug screening, diagnostic imaging of cancer samples, and live imaging of cancer cell dynamics. Yet one limitation has largely remained; the low number (~4) of molecular markers observable in each cell without recourse to complex/expensive sequential or spectral multiplexing methods. Though deep learning-based virtual labelling of cells has thus far provided a nascent alternative to predict a few markers per cell, more extensive virtual multiplexing has yet to be explored, including to empower downstream phenotypic analyses.
Design/Methods
Based on standard 4-channel immunofluorescence (IF) of A549 lung cancer cells during epithelial-mesenchymal transition (EMT), we used a novel combination of label-free and fluorescence inputs with a state-of-the-art transformer to predict EMT marker labelling in thousands of single cells. Cells treated with control, EGF or TGF-β had phenotypes spanning the EMT spectrum and we leveraged this heterogeneity to show that virtual multiplexing significantly improves phenotype classification. Unsupervised PHATE manifold embedding also emphasised how virtual multiplexing improves resolution of single cell states across the EMT spectrum.
Results
Phenotyping of virtually multiplexed cells matched/outperformed standard 4-channel IF, enabling precise EMT-treatment classification. Manifolds formed using virtually multiplexed cells recapitulated known inter-marker relationships inaccessible in 4-channel data, reinforced by observable established morphological transitions over the EMT manifold. Compared to the unstructured manifold based on standard 4-channel IF, we show for the first time the ability of virtual multiplexing to reveal structures of cellular heterogeneity beyond those resolvable using standard IF.
Conclusion
We show how virtual multiplexing provides single-cell analyses with phenotypic resolution far exceeding that of standard IF. Using a novel combinatorial approach linking label-free and standard IF inputs, we also uniquely show that vision transformers outperform previous models for virtual labelling. Now facilitating enriched single cell phenotyping in innumerable research settings – without demand for complex multiplexing capabilities – our novel approach also permits mapping of inter-marker relationships and detailed cell states otherwise inaccessible without far more complex experimental systems.