Fig. 4: ExIF improves phenotypic classification of experimental treatments modulating epithelial-mesenchymal state in comparison to standard IF.

A Workflow comparing cell phenotype classification using different real/virtual label combinations. Fields imaged for DIC, 3 ‘general markers’ (GM) and a single epithelial-mesenchymal marker (EM) (described in Fig. 3) and used for virtual labelling and for cell segmentation/feature extraction. Combinations of features from real/virtual markers underpinned support vector machine (SVM) classification of phenotypes from control, EGF, and TGF-β1 conditions. Microscope icon by DBCLS licensed under CC-BY 4.0 Unported. B 5-fold cross-validated classification accuracy (F1-score) using: different virtual labelling inputs (table, left column); features from virtual markers (table, middle column); and features from real images (table, right column). General markers, single variable marker, and all 8 variable markers respectively denoted by 3GM, 1EM, 8EM. Box plots depict data distributions as follows: dashed centre line, median; solid centre line, mean; box limits, upper and lower quartiles; whiskers, 1.5x interquartile range. Significance reflects Welsh’s two-sided t-testing. C Confusion matrices for classification using real general markers (left; as per first row in B); or Full ExIF (centre, as per last row in (B)). Differences in performance across these confusion matrices (right). D Example control cells (green, purple) misclassified as TGF-β1-treated when using general markers (as per first row in (B)) and corrected by hybrid multiplexing using real general markers and 8 virtual EM-state markers (as per last row in (B)) (markers as labelled combined in grey). E Comparison of real (solid boxes) or virtual (dashed boxes) mean intensity distributions (normalised histograms) for cells from control (blue) or TGF-β1-treated (red) conditions. Mean intensity values of the control cells denoted (vertical lines as per (D)). For these cells, real general marker intensities provide little distinction between treatments. By contrast, several virtual EM labels strongly distinguish conditions, correcting classification of the exemplar cells. Left images show real/virtual labels from the exemplar field (from (D)), cropped images (centre) show exemplar cells. For comparison, right images show an exemplary TGF-β1-treated field for the same indicated real/virtual marker. Source data provided as detailed in data availability.