Seeing more with less: Extensible Immunofluorescence (ExIF) accessibly generates high-plexity datasets by integrating standard 4-plex imaging data

doi:10.21203/rs.3.rs-4872456/v1

Download PDF

Article

Seeing more with less: Extensible Immunofluorescence (ExIF) accessibly generates high-plexity datasets by integrating standard 4-plex imaging data

https://doi.org/10.21203/rs.3.rs-4872456/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Standard immunofluorescence imaging captures just ~4 molecular markers (‘4-plex’) per cell, limiting dissection of complex biology. Inspired by multimodal omics-based data integration approaches, we propose an Extensible Immunofluorescence (ExIF) framework that transforms carefully designed but easily produced panels of 4-plex immunofluorescence into a unified dataset with theoretically unlimited marker plexity, using generative deep learning-based virtual labelling. ExIF enables integrated analyses of complex cell biology, exemplified here through interrogation of the epithelial-mesenchymal transition (EMT), driving significant improvements in downstream quantitative analyses usually reserved for omics data, including: classification of cell phenotypes; manifold learning of cell phenotype heterogeneity, and; pseudotemporal inference of molecular marker dynamics. Introducing data integration concepts from omics to microscopy, ExIF provides a blueprint empowering life scientists to use routine 4-plex immunofluorescence methods to achieve previously inaccessible high-plex imaging-based quantitative single-cell analyses.

Biological sciences/Computational biology and bioinformatics/Machine learning

Biological sciences/Computational biology and bioinformatics/Software

Biological sciences/Computational biology and bioinformatics/Image processing

Biological sciences/Biological techniques/Imaging/Fluorescence imaging

Biological sciences/Computational biology and bioinformatics/Data integration

Multiplex

multi-modal data integration

generative deep learning

transformer

single-cell analysis

epithelial-mesenchymal transition

feature engineering

Heterogeneous cell states and phenotypes arise from complex processes shaped by the levels and subcellular localisations of various types of molecular component, including DNA, RNA, proteins, metabolites and lipids. Over the past decade, many omics modalities have emerged to systematically define molecular component levels with single-cell resolution for (typically) one of these component types at a time across heterogeneous cell populations^1,2. More recently, computational methods have been developed to integrate different molecular component profiles captured via multiple omics modalities^3,4. As recently reviewed⁵, the most effective of these data integration strategies leverage so-called data ‘anchors’, i.e. measured features and/or cell populations that are common across otherwise independent datasets and which guide the quantitative integration process. Particularly in the presence of such data anchors, ‘multimodal’ data integration can dramatically enrich analyses of complex multimolecular cell states and mechanisms.

Compared to existing omics technologies, fluorescence microscopic imaging remains unparalleled in its ability to provide measurement resolutions (especially sub-cellular spatial and temporal resolution) exceeding those accessible via omics⁶. Yet, despite these advantages in resolving power, almost no emphasis has yet been placed on developing computational approaches that can integrate fluorescence imaging data from distinct markers and cell populations to achieve advances in molecular sampling depth and breadth analogous to those in omics fields. Such advances are however needed, because the majority of fluorescence imaging practitioners are limited to concurrent observation of just ~4 molecular components per cell/sample, due to long-standing technical constraints⁷ (e.g. spectral bleed-through, lack of antibody species diversity). This low-plexity limitation remains pervasive despite the emergence of powerful experimental methods enabling high-plexity immunofluorescence (IF) imaging^8–10, because the accessibility of these experimental methods is severely restricted by their complexity, cost and time-demands. Thus, most researchers leveraging the enhanced resolving power of fluorescence imaging methods remain inhibited in their ability to interrogate complex, multimolecular cell processes.

To democratise a solution for this pervasive limitation, we here take inspiration from the field of multimodal omics data integration to develop ‘Extensible Immunofluorescence’ (ExIF) (Figure 1). ExIF comprises a mixed experimental and computational framework for integrating a theoretically unlimited number of molecular markers (plexity) in distinct single-cell populations without the need for specialised (e.g. spectral) hardware or potentially degradative reagents¹¹. Instead, an array of different standard 4-plex IF panels (each labelling an independent cell population, e.g. in different wells of a multi-well plate) are designed to permit ‘extensible labelling’ (Figure 1A) wherein each 4-plex panel contains a mixture of one or more ‘anchoring channels’, which recur in every panel (i.e. across all cell populations), and one or more ‘variable channels’, which differ in each additional labelling panel. Such extensible labelling initially gives rise to non-integrated data (Figure 1B) comprising multiple 4-plex IF image sets spanning various channels and cell populations. The anchoring channels are then used as constant inputs for generative deep learning (DL) models that are trained to produce ‘virtual labelling’^12,13 of eachvariable channel target (Figure 1C). Since anchoring channels are (by design) present in all cell populations, all virtual labelling models (one per variable channel-target) can also be applied in all cell populations, thus integrating every variable channel marker into every cell. ExIF thereby creates a unified data matrix (Figure 1D) combining all molecular marker channels across all cells, using high-fidelity virtual labelling to mediate a specialised form of mosaic integration⁵ that we term 'bridging' integration (reflecting the use of designed data anchors that support DL models to span and connect the dataset).

Significantly, to date, virtual labelling has overwhelmingly been utilised within a paradigm of ‘label-replacement’, i.e. to obviate the need for fluorescence labelling by training DL models to predict a fluorescence marker-of-interest using a label-free input^12–24. Whilst powerful in specific contexts, these approaches have until now fallen short of a structured integration framework permitting scalable increases in both molecular marker plexity per cell and total sampled cell population. Here, by shifting from a label-replacement model to one of dataset-integration, we demonstrate the capacity for virtual labelling to generate large-scale, high-plexity datasets, requiring only a single experiment for both DL model training and downstream biological analyses through ExIF.

In this study, after first optimising DL model architectures and input options to enhance virtual labelling fidelity beyond the state-of-the-art, we demonstrate the utility of ExIF by showing how a simple implementation can dramatically enrich downstream single-cell quantitative analyses. Specifically, in the context of epithelial-mesenchymal state heterogeneity, we show: i) improved machine learning (ML)-based classification of cell state responses to EM-transition (EMT)-modulators (epidermal growth factor, EGF; transforming growth factor-beta one, TGF-β1); ii) enhanced manifold learning to map single cell heterogeneity in EM-phenotypes; and iii) more precise inference of EM-state marker dynamics during EMT. Excitingly, in each case, ExIF-integrated datasets far exceed the performance of the standard 4-plex IF datasets from which they are derived. Overall, ExIF establishes a new methodological framework for the computational integration of fluorescence imaging data, democratising capacity for high-plexity, high-resolution imaging-based interrogations of complex, multimolecular cell biology.

Single-cell integration through virtual labelling

Generative DL-enabled virtual labelling mediates the integration of molecular marker channels and cell populations within our ExIF framework (Figure 1). Since virtual labelling fidelity thus determines the quality of ExIF data integration, we begin this study by optimising virtual labelling performance, first by comparing alternate generative DL architectures, and then by testing different types and combinations of anchoring channels (DL model inputs).

Reflecting the prevailing ‘label-replacement’ paradigm of virtual labelling^12,15, we first use label-free differential interference contrast (DIC) images as input (Figure 2A, left) to compare three alternate generative DL architectures: U-Net²⁵; cGAN²⁶; and ResViT²⁷- the latter of which incorporates self-attention and is used here for the first time in the context of virtual labelling. We compare virtual labelling performance across eight alternate molecular marker channels (α-Tubulin, DNA (DAPI), CoxIV, fibrillarin, GM130, F-actin (phalloidin), β-Catenin, NF-κB-p65) concurrently labelled via the ‘4i’ experimental multiplexing protocol²⁸ in 45,780 DU145 prostate cancer cells (Figure 2B). We stress that (experimentally) multiplexed IF datasets are not required for ExIF, which is explicitly designed to enable high-plexity dataset integration based on standard 4-plex IF data (as demonstrated across Figures 3-5). Rather, the experimentally multiplexed dataset was used here for developing and testing computational elements of the ExIF framework. All comparative results reported here and throughout this study are based on robust 5-fold cross-validation.

Virtual labelling fidelity with DIC-only inputs was first assessed using image-level metrics: the structured similarity index (SSIM²⁹) and Pearson’s Correlation Coefficient³⁰ at the pixel level (Pixel-PCC) by comparing predicted virtual labels and their real counterparts. U-Net achieved the highest overall SSIM scores, while the ResViT model achieved the highest Pixel-PCC scores (Figure 2C). To focus on virtual labelling fidelity at single-cell resolution, we next performed cell segmentation (Supplementary Figure 1A) and marker feature measurement using CellProfiler³¹ software, comparing 142 quantitative features per marker channel, per cell, spanning several feature categories: 15 ‘intensity’; 75 ‘radial distribution’, and 52 ‘texture’ features (complete feature list in Supplementary Table 1). For each quantitative feature, we calculated the Pearson’s correlation (metric herein termed ‘Feature-PCC’) between ground-truth experimentally labelled and virtually labelled marker channels. Significantly higher Feature-PCC values were achieved across all feature categories using ResViT (Figure 2D). For per cell mean marker intensities – particularly important given their relation to protein abundance – ResViT predictions achieved high correlations to ground-truth values for most markers (~0.6 - 0.9 PCC; Figure 2E), though GM130 and β-Catenin correlations were modest (<0.5 PCC). A full breakdown per marker, DL architecture and image metric/feature category is provided in Supplementary Figure 1B. Given the convincing outperformance of ResViT, particularly at the per-cell feature level, we hereafter present ResViT results only, together with Feature-PCC metrics of per cell quantitative feature similarity between real and virtual labels, as these emphasise biologically interpretable analyses at single-cell scale.

We next targeted further improvements in virtual labelling fidelity by providing ResViT with multiple input image channels. We initially complemented the DIC anchoring channel with a single additional fluorescence marker channel (Figure 2A, center) (testing all alternatives) and found improvements in Feature-PCC scores in essentially all cases, although improvement levels varied for mean intensity (Figure 2F) as well as across each feature category (Supplementary Figure 2A). Unsurprisingly, the choice of additional input image channel made a marked difference to the degree of improvement observed, as exemplified when integrating β-Catenin or GM130 virtual labels across the cells (Figure 2G, Supplementary Figure 2B). Extending this multichannel input strategy to its limit, we built ResViT models incorporating DIC plus seven fluorescence marker channels as input to predict an eighth marker (Figure 2A, right). Feature-PCC values improved further across all target markers, though increases became more subtle beyond the impact of adding the single best fluorescence marker to DIC, when considering mean intensity prediction (Figure 2F) or other feature categories (Supplementary Figure 2A). Taken together, we have achieved significant increases in virtual labelling fidelity through innovative use of the ResViT DL architecture, and by providing ResViT with multiple image inputs as anchoring channels.

Applying extensible immunofluorescence to investigate epithelial-mesenchymal cell state plasticity

We next demonstrate ExIF in the context of epithelial-mesenchymal plasticity, a phenomenon facilitating cancer progression by enabling tumour cells to switch between epithelial and mesenchymal phenotypes^32,33. Epithelial-mesenchymal transition (EMT) is a complex multimolecular process, involving repression of epithelial markers and upregulation of mesenchymal markers enabling enhanced cell migration, invasion and thus metastatic progression³⁴. To assess the utility of ExIF for interrogating EM cell state plasticity, we amplified EM state diversity in A549 lung cancer cells by comparing control cells to those treated for 48 h with either EGF (0.1μg/mL) or TGF-β1 (0.01μg/mL); growth factors known to weakly or strongly (respectively) drive epithelial-to-mesenchymal transition (EMT) in A549 cells³⁵. To readout EM state per cell, we labelled and imaged eight different 4-plex IF panels across each treatment condition, following the extensible labelling schema (Figure 1). Specifically, each panel consists of recurring ‘general’ markers (DNA, F-Actin and β-Catenin), and a single variable EM state marker (E-Cadherin or EpCAM or N-Cadherin or PTEN or Vimentin or CD44total or CD44std or CD44v9^36–41(Figure 3A).

We then performed data integration via virtual labelling, using the recurring general markers and DIC as inputs to ResViT virtual labelling models to predict labelling for variable (non-recurring) EM state markers. We highlight the role played by the recurring markers in the ExIF context, providing consistent virtual labelling inputs and acting as common anchors for accurate integration of variable markers. This enables production of a unified ExIF dataset, with each cell characterised by all 11 fluorescence markers (3 general markers and 8 EM state markers). CellProfiler-based cell segmentation and feature measurement followed (morphological features (cell size, shape, neighbour-contact etc) plus 142 features per marker per cell; as in Supplementary Table 1), acquiring data spanning 12,103 control cells, 14,350 EGF-treated cells and 7,119 TGF-β1-treated cells.

Although ExIF is designed to enable utilisation of multiple input channels to improve integration performance, we note that some research applications are suited to a label-free only approach (e.g. long-term live cell imaging, sensitive diagnostic sample imaging etc). Thus, in addition to assessing integration performance using ‘full ExIF’ (using general fluorescence markers and DIC as input to predict each of the 8 EM-state markers), we also assessed using DIC as the only input; termed ‘label-free ExIF’ (Figure 3B). Consistent with our previous results, we found significantly higher Feature-PCC values between real and virtual labels when using full-ExIF (Figure 3C; Feature-PCC values per feature category, per marker shown in Supplementary Figure 3A). The multichannel inputs achieved extremely high correlations (~0.7 – 0.9 PCC) for per cell mean intensity of virtual labels relative to real labels (though vimentin showed weaker performance; < 0.6 PCC) (Figure 3D).

Using EM state marker mean intensity distributions to examine differences induced by growth factor treatments (Figure 3E), we found that control and EGF conditions had relatively similar distributions, whereas TGF-β1 induced more dramatic changes. Mesenchymal markers N-Cadherin and Vimentin were especially strongly expressed after TGF-β1 treatment, while epithelial markers CD44v9, E-Cadherin, and EpCAM were strongly suppressed. These effects are consistent with established EMT biology^36–41. Notably, EM marker distributions resulting from virtual labelling-mediated dataset integration closely reflected treatment-induced effects seen with real labels (i.e. those derived from the subset of cells experimentally labelled with each EM marker). Thus, ExIF achieves strong performance across various EM state markers, retaining biologically relevant information that captures heterogeneous yet distinctive responses to different EM-state perturbations.

Classifying cell treatment condition using ExIF integrated data

We next assessed whether ExIF meaningfully enhances downstream single-cell quantitative data analyses commonly used to interrogate complex cell biology, including in the contexts of omics and multi-omics analyses. We first tested accuracies for machine learning classification of treatment conditions (control vs EGF vs TGF-β1). Using standardized and scaled single-cell (CellProfiler-derived) features from standard and ExIF integrated datasets, we performed principal components analysis (PCA) before using derived principal components as input for support-vector machine (SVM)-based classification of cell treatment conditions (Figure 4A). This was repeated for each cell population labelled with a different EM state marker. Classification results were evaluated via F1-scores, comparing different datasets in aggregate (Figure 4B).

We found that full ExIF achieved significantly better classification of treatment conditions (F1 ~ 0.85) compared to standard 4-plex IF (F1 ~ 0.8). We also noted that ExIF anchored using only the three general fluorescence channels (no DIC) retained classification performance (F1 ~ 0.84) that was far better than when using only features directly measured from the three general fluorescence anchoring channels (F1 < 0.75). This implies that ExIF integration of EM-state markers (via virtual labelling) adds significant discriminatory value (regarding treatment class) beyond that directly accessed from the general markers that form the basis of integration. Label-free ExIF, which virtually labels all 11 fluorescence markers from this experiment also significantly improved classification beyond that achieved using the three (real) general markers. Indeed, label-free ExIF achieves essentially the same EM-state discriminatory power with label-free images-only as can be achieved using features from a standard 4-plex IF experiment that includes at least one explicit EM-state marker. By comparison, we note that virtual labelling of only the original 4-plex panels using DIC-only-inputs – an approach that mimics the prevailing label-replacement paradigm for virtual labelling – failed to match even classification results achieved using features measured from the three general markers.

Interestingly, classification using only features from the three general markers most often misclassified control and TGF-β1 cells as EGF-treated (Figure 4C), while full ExIF integration reduced misclassifications across all classes, especially improving differentiation of TGF-β1 and EGF conditions. To explain these performance-differences, we show examples of how the addition of virtual markers corrects two control cells misclassified as TGF-β1-treated (Figure 4D). Comparing EM-state marker mean intensity distributions, for instance, we see no clear separation between control (blue) and TGF-β1 conditions (red) when considering the real common anchoring channels-only, although TGF-β1-treated cells have somewhat lower mean DNA intensities (Figure 4E). Given these minimal general marker differences between control and TGF-β1 conditions, the two misclassified control cells (mean intensities marked by vertical lines - green or purple) could belong to either treatment population. Indeed, the real general marker labelling of these control cells looks similar to exemplar TGF-β1-treated cells. In contrast, EM-state marker integration, especially of CD44v9, E-Cadherin, N-Cadherin, EpCAM, and Vimentin, defines population distributions that more effectively discern control and TGF-β1 conditions. Accordingly, the two example cells can be identified as strongly exhibiting control phenotypes when considering the combination of all virtual markers, exemplifying how ExIF integration enhances phenotype discrimination and classification capacity.

Mapping EM-state heterogeneity and marker dynamics using ExIF data integration

Moving beyond classification of the discrete treatment conditions, we next examined whether ExIF integration could enhance the quantitative mapping of cell phenotype heterogeneity in the EM spectrum. To this end, we applied several unsupervised manifold learning methods to our quantitative single-cell feature datasets. Importantly, we excluded cell morphology features (relating to cell size, cell shape, cell neighbour proximity) from construction of these and all subsequent manifolds, since cell morphology is linked to EM state, and thus morphological features would undermine capacity to assess the construction of ‘EM-state-sensitive’ manifolds leveraging ExIF-integrated data. We present manifolds constructed with PHATE, a manifold-embedding technique that seeks to capture underlying dynamic trajectories in multivariate data⁴² (Figure 5), as well as manifolds constructed with T-SNE⁴³ and UMAP⁴⁴ (Supplementary Figure 3B).

For the ‘Standard’ manifolds, we used features measured directly from the three general markers only (Figure 5A), demonstrating the limited capacity of standard-plexity IF data to map heterogeneity in EM-state. We next constructed manifolds via label-free ExIF integration of all 11 markers using DIC as the only DL model input (Figure 5B). We found that the label-free ExIF integrated dataset maps EM-state heterogeneity more effectively than standard IF data, with improved treatment condition-separation and manifold extension. Lastly, we generated full ExIF manifolds (using general markers plus DIC as DL model inputs), integrating all eight EM-state markers (Figure 5C). Features measured from this optimally integrated dataset greatly improved delineation of treatment conditions in the PHATE manifold, emphasising differences between the ‘mesenchymal’ states induced by EGF versus TGF- β1^45–47, as well as defining a more structured and extended manifold profile than the other manifold versions. T-SNE and UMAP manifolds confirm similar tendencies.

Modelling of pseudotime⁴⁸ within the PHATE-embedded manifolds further supports improved cogency of the full ExIF manifold, which defines a trajectory capturing progression from cell states enriched for the control condition to cell states enriched for EGF then TGF-β1 treatments. Representative cell images from milestones (1-6) along the full ExIF manifold pseudotime trajectory indicate cell morphological trends aligned with EMT (Figure 5D). Indeed, colour-coding the full ExIF manifold by real (CellProfiler-measured) morphological and cell-cell contact values (not used in manifold construction) shows consistent quantitative trends aligned with expected EMT dynamics, i.e. increasing cell area, increasing cell protrusivity and reduced cell-cell contact (Figure 5E). Pseudotime analyses further delineated individual EM marker-level (mean cell intensity) variations along each inferred trajectory. These variations are unstructured and noisy in the Standard manifold compared to clear trends and signals spanning the full ExIF manifold (Figure 5F); again supporting improved manifold characteristics achieved via optimal ExIF dataset integration. Notably, the displayed EM marker dynamics derived from the Standard manifold (Figure 5F, left) reflect real marker values from the cell subpopulations experimentally labelled for each variable marker. EM marker dynamics from the full ExIF manifold (Figure 5F, right) depict either: real marker dynamics (weak line colours; from the subset of cells experimentally labelled for each marker); or virtual marker dynamics (strong line colours; from all cells). Comparisons of real versus virtual marker trends reveal strong correspondence for each marker, indicating that dynamical trends inferred using the virtual data used for dataset integration correspond to ground-truth marker dynamics, though the virtual data trends are substantially less noisy.

Focusing on the specific expression-level dynamics inferred for each EM marker in the optimally performing full ExIF manifold, we see trends aligned with established molecular changes during transition from epithelial to mesenchymal state. E-cadherin expression declines while N-cadherin levels ultimately elevate⁴⁹. The epithelially-associated CD44v9 isoform is suppressed, counterbalanced by increases in total CD44 and in the mesenchymally-associated CD44std isoform⁵⁰. Tumour-suppressor PTEN expression declines while vimentin levels peak late in the trajectory^39,40. As with E-cadherin, EpCAM also declines substantially⁵¹. By contrast, these biologically informative trends are difficult or impossible to discern from pseudotime analysis of the Standard manifold based on data from the common fluorescence anchoring channels-alone.

Taken together, we find that ExIF integration – even when using only label-free (DIC) inputs – can achieve analytical outcomes that compare favourably to interrogation of standard IF data alone. Moreover, our full ExIF approach, which achieves high-fidelity integration to create high-plexity fluorescence imaging data, enables powerful multimolecular analyses of discrete cell states, continuous phenotypic heterogeneity and (pseudotemporal) marker dynamics. Achievable with just a single experiment comprising several (or many more) standard 4-plex IF labelling conditions, these results are comparable to those otherwise accessible only by using complex and costly experimental multiplexing methodologies.

Compared to the single-cell omics domain^1,2 where computational methods for data integration have rapidly proliferated, equivalent computational integration methods remain entirely lacking for fluorescence imaging-based research applications, where the benefits of heightened resolution⁶ have long been counterbalanced in the context of multimolecular biological analyses by severe restrictions on molecular marker plexity. Though several experimental methods permitting high-plexity fluorescence imaging have emerged^10,28,52, their uptake is low due to significant complexity, cost and time-consumption, thus emphasising the need for more accessible strategies. Accordingly, to democratise capacity for high-plexity fluorescence imaging analyses we have developed the Extensible Immunofluorescence (ExIF) framework, combining concepts from multimodal omics data integration with innovations in generative deep learning-enabled virtual labelling. ExIF unifies carefully designed ‘extensible labelling’ panels of otherwise routine 4-plex IF images into theoretically unlimited-plexity datasets using virtual labelling to mediate computational data integration. We have shown that ExIF integrates cells and markers with high fidelity to significantly enrich downstream biological interrogation of both discrete (classification) and continuous (manifold learning) cell phenotype heterogeneity, as well as improved inference of EM-state marker dynamics, achieving analytical capacity not afforded by standard IF alone. ExIF thus makes high-resolution fluorescence imaging-based multimolecular analyses far more accessible to the life science research and translation community.

Though until now lacking translation to the domain of fluorescence imaging, growing interest in integration for omics data^3,4 has led to a conceptual schema defining several types of data integration strategy. These include so-called ‘diagonal’, ‘horizontal’, ‘vertical’ and ‘mosaic’ integration strategies that are defined by the presence and nature of the data ‘anchors’ (recurring cell populations and/or molecular features) that serve to connect disparate datasets⁵. Briefly, diagonal integration is necessary (though challenging) where no common data anchors link independent datasets, whereas common molecular features (horizontal integration; i.e. equivalent feature-sets measured across multiple cell populations) or cell populations (vertical integration; i.e. multiple feature-sets measured in one cell population) anchor integration in other contexts. Finally, mosaic integration relies on partial overlap (often random) between cell populations and/or measured features enabling data-driven connections between all data subsets, wherein any pair of data subsets may be linked by different anchors. Since ExIF performs both cell-wise and feature-wise integration, it is akin to mosaic integration, yet it differs in that our novel (though easily reproduced) extensible labelling strategy allows us to design the anchoring features that are common across cell populations, giving opportunities to optimise their integrative performance while also removing the random element of overlap that can undermine mosaic integration performance. Given the structured nature of the anchors that support DL models to connect across the total span of the dataset, we refer to ExIF as a ‘bridging’ integration strategy. This novel ‘bridging’ approach exemplifies how careful experimental design can aid in the formulation of systematic integration approaches, and for the first time brings a data integration strategy to the domain of fluorescence imaging.

The use of virtual labelling in ExIF to integrate disparate cells and markers represents a significant advance in the framing of virtual labelling utility, where to date, emphasis has overwhelmingly been on replacing real labels with virtual labels to maintain specific molecular insights in the absence of experimental labelling^12–23 – a mode of utilisation that we refer to as ‘label-replacement’. Given this prevailing label-replacement paradigm, it is unsurprising that the use of fluorescence inputs (as opposed to label-free inputs) for virtual labelling models has thus far received little attention⁵³. More surprising has been the relatively limited emphasis on assessing whether virtual labelling can enhance downstream quantitative biological analyses^53,54. By contrast, the ExIF framework now utilises virtual labelling to achieve data integration, which, combined with our extensible 4-plex IF labelling approach, can unify high-plexity, high cell number datasets for rigorous quantitative interrogation of complex cellular processes. Resulting ExIF-mediated improvements in analytical performance are particularly noteworthy in several cases. In our analysis of EM states, comparing treatment classification performance between the ‘general markers’ (3GM) condition and the ‘fluorescence ExIF’ condition (which uses the same 3 general markers as anchors for data integration) reveals a remarkable increase in classification accuracy due to virtual labelling-enabled data integration. Crucially, this is despite both analyses ultimately relying on identical underlying image data (the 3 general markers). Similar analytical performance improvements were observed when comparing the standard manifold to the full ExIF manifold – with the major differences in input to these manifolds being the integration of all 8 virtual EM-state markers (into the full ExIF manifold). Together, these results suggest that data integration through the virtual labelling process functions as a form of feature engineering⁵⁵ that surfaces latent information in the anchoring channels and makes this more accessible to the finite machine learning models used for either classification (SVM) or manifold learning (PHATE, t-SNE, UMAP) – thereby improving their performance.

When considering the design of experiments employing the ExIF framework, we note differences in virtual labelling performance (and thus dataset integration) depending on the selection of anchoring markers. While we have chosen morphological (DNA, F-actin) and EM-state adjacent (b-catenin) markers to act as anchors, our results suggest there is potential to optimise anchoring channel selection to improve the fidelity of targeted virtual labels. At the same time, we also saw that ExIF integration anchored by the label-free (DIC) channel only achieved analytical performance matching that of standard 4-plex IF. This indicates significant potential for such high-plex virtual label integration to enhance cell phenotype analyses even in purely label-free imaging applications such as ultra-high-throughput screening, long-term live cell imaging, or minimally invasive imaging of sensitive clinical samples⁵⁶.

In summary, we have combined an original yet simple extensible labelling experimental design with innovations in the implementation and application of virtual labelling to establish the Extensible Immunofluorescence (ExIF) framework. Using standard 4-plex IF, ExIF enables simultaneous integration of both cells and molecular markers to generate unified datasets with high-fidelity to experimental labelling. In comparison to standard IF, we have demonstrated major improvements in capacity to perform several downstream quantitative single-cell, multimolecular analyses. Thus, the ExIF framework achieves a step-change in the utility of standard IF as a method for quantitative single-cell analyses of complex, multimolecular biology. This democratises the powerful benefits of high-plexity fluorescence imaging for the thousands of laboratories within which standard IF methods are already routine.

Cell culture and imaging

Cell culture

DU145 and A549 cells were cultured in Minimum Essential Medium (MEM, Thermo Fisher Scientific) and Dulbecco's Modified Eagle Medium (DMEM, Thermo Fisher Scientific), respectively. Both mediums were supplemented with 10% Fetal Bovine Serum (FBS, Thermo Fisher Scientific), 4mM L-glutamine (Thermo Fisher Scientific) and 23.4 mM 4-(2-hydroxyethyl)-1-piperazineethanesulfonic acid (HEPES, Thermo Fisher Scientific).

Cells were passaged two to three times per week and used within passage number 3 and 20. Cells were cultured at 37˚C with 5% CO₂ and 21% O₂to 85% confluence before harvesting with 1x DPBS and 0.05% Trypsin-EDTA (Thermo Fisher Scientific) incubation for 3-5 min.

EMT activation

For EMT activation, 3,000 A549 cells were seeded into 96-well glass bottom plates (Cellvis). After 20 hrs, DMEM was removed, and cells were cultured with serum-free media for 24 hrs. Cells were then treated for 48 hrs with either 0.1µg/mL human epidermal growth factor (hEGF, Cell Signalling Technology, #8916) or 0.01µg/mL transforming growth factor beta 1 (TGF-β1, R&D Systems, #5036-WN) diluted in serum-free media.

Immunofluorescence and Iterative indirect immunofluorescence imaging (4i)

For 4i multiplexing experiments, 10,000 DU145 cells were seeded into 96-well glass bottom plates (Cellvis) and 4i was performed as previously described^28,57. High-quality antibodies, previously validated for use in 4i, were chosen for labelling and are listed in Supplementary Table 2⁵⁸. Successful antibody elution was assessed at each cycle of the 4i process. This evaluation was performed by imaging elution control wells, which consisted of re-probing previously labelled, imaged, and eluted wells with secondary antibodies. Complete elution was verified for all cycles.

For analysis of EMT responses, A549 cells were fixed with 4% paraformaldehyde (Electron Microscopy Sciences) for 15 min and permeabilised with 0.1% Triton X-100 (Thermo Fisher Scientific) for 20 min at room temperature (RT). Cells were blocked for 1 hour with TBS intercept blocking buffer (LI-COR Biosciences) before being incubated with primary antibodies diluted in TBS intercept blocking buffer for 2 hours at RT. Cells were then incubated with secondary antibodies diluted in TBS intercept blocking buffer for 30 minutes at RT. All antibodies are listed in Supplementary Table 2. Cells were washed three times with TBS after fixation and permeabilization, as well as after the primary and secondary antibody incubations.

Confocal microscopy

Images were acquired with a Nikon A1R confocal using JOBS automation for reproducible, unbiased multi-position, multi-well acquisition. Automated laser-based and image-based (using DAPI labelling) focusing were combined to ensure optimal, reproducible focusing. A 20x Plan Apochromat air objective (NA 0.75) was used to enable scalable imaging at high-speed via an HD25 resonance scanner at 1024 by 1024-pixel resolution. Pixel saturation was avoided through monitoring of high-low look up tables during laser power and gain setting to preclude pixel value clipping and resulting value ambiguity in the training of deep learning models. Raw 16-bit .nd2 image files were converted to TIFF format using custom Python scripts prior to quantitative analyses. Brightness and contrast were optimized with Fiji software (National Institutes of Health) for visualisation and presentation purposes only.

Image analysis, virtual labelling, and downstream quantitative analyses

Pre-processing

Background removal was performed on label-free images by subtracting the median (40px kernel) filtered image in the experimentally multiplexed (A549 cell) dataset and subtracting a Gaussian blurred image (σ = 0.5) for the EMT (DU145 cell) dataset due to the different imaging artefacts in each dataset. Contrast enhancement was then applied, allowing 35% of the pixels to become saturated. Images were then converted to 8-bit/pixel by performing linear scaling, projecting the min-max to 0-255 respectively in Fiji software. Fluorescence images were converted to 8-bit by performing linear scaling with respect to the whole dataset. The top and bottom 10% of pixels across all fluorescence images were treated as outliers and the min-max of remaining pixels in the dataset were projected to 0-255 respectively (saturating the outliers). All images were then resized by half using bilinear interpolation, before being cropped into patches measuring 256x256. This resulted in a total of 583 patches of DU145 cells, each with – DIC, α-Tubulin, DNA, CoxIV, Fibrillarin, GM130, F-Actin, β-Catenin, and NF-κB channels; and 888 patches of A549 cells, each with DIC, DNA, F-Actin, and β-catenin as well as one of the following variable markers: E-Cadherin, EpCAM, N-Cadherin, PTEN, Vimentin, CD44total, CD44std, CD44v9.

Virtual labelling model design and training

Three models were considered for virtual training. First, a model based on U-Net²⁵ was used, consisting of down-convolution blocks with skip connections to up-convolution blocks. Convolutions were used with a 4x4 kernel, 2x2 stride and 1x1 padding, followed by ReLU activation (leaky for down-convolutions with 0.2 negative slope) and then batch normalisation (ε = 10^-5, momentum = 0.1) apart from the first and last convolutional layers). Layer size was initialised at 64 filters and doubled until 512 filters. A total of 8 down-convolution layers were used with corresponding 8 up-convolution layers; deeper than the original U-Net as we informally found small improvements with increasing depth. 0.5 Dropout regularisation was added for 3 of the layers before the deepest layer. Tanh activation was used after the final up-convolution. L1 loss was used as the objective function.

Second, the cGAN model²⁶ utilised the U-Net backbone described as a generator network and was extended with adversarial training. The discriminator network was comprised of down-convolutions with a 4x4 kernel, 2x2 stride and 1x1 padding starting with 64 filters and doubling until 512 filters. Each convolution layer was followed by leaky ReLU activation (0.2 negative slope), and batch normalisation (ε = 10^-5, momentum = 0.1) apart from the first layer. The final down-convolution layer had 1x1 stride. A final classification layer was also added with 1x1 stride. We combined L1 loss with LSGAN loss which is the mean squared error of the discriminator’s prediction.

Both U-Net and GAN models were trained for 50 training epochs with a learning rate of 0.0002 before linearly decaying for another 50 epochs as conducted by others²⁷.

Finally, a residual vision transformer model – Residual vision transformer (ResViT) was used²⁷. It consists of an encoder and decoder module consisting of CNN blocks, and a bottleneck in between composed of a series of aggregated residual transformer (ART) blocks composed of multi-head self-attention and convolutions. The network (without the transformer) pretrained on the public ImageNet21k dataset and was further trained for 50 epochs with a learning rate of 0.0002 before linearly decaying for another 50 epochs. The whole network (including the transformer) was then trained for another 25 epochs with a learning rate of 0.0002 before linearly decaying for another 25 epochs²⁷.

Optimisation was done using ADAM (β₁ = 0.5, β₂= 0.999) for all three models⁵⁹.

Model training was conducted using 5-fold cross-validation, stratified with respect to treatment conditions for EMT-activated cells. Data augmentation by vertical and/or horizontal flipping expanded the number of unique images by a factor of four. To normalise the data, image pixel intensities were scaled between [-0.5, 0.5]. Model training was conducted using an NVIDIA V100 GPU taking roughly 4 hours per fold.

Segmentation and Feature extraction

For downstream analysis and performance evaluation, features were extracted per cell using a custom workflow in CellProfiler⁶⁰. First, cell segmentation was performed using the default ‘Cells’ Cellpose model on experimental DNA (DAPI) and F-Actin (Phalloidin) channels (Supplementary Figure 1A). Objects in contact with the image border were excluded to remove partial cells. In total, per cell, we extracted pairwise marker signal colocalization (Pearson’s correlation coefficient based on per pixel intensities) plus 15 intensity, 75 radial distribution, and 52 texture features (see Supplementary Table 1). For the 8-plex experimentally multiplexed DU145 dataset (section 1), 45,780 cells were assessed from a single biological replicate with 8 technical replicates (wells). For the EMT dataset based on A549 cells (section 2), 33,572 cells were analysed from a single biological replicate across 24 distinct treatment and labelling conditions (wells).

Evaluation

Predicted labels were evaluated at the image-level using the structural similarity index measure⁶¹ (SSIM) for a 7x7 window and Pearson’s correlation coefficient between corresponding pixels (Pixel-PCC). Co-localisation of predicted and experimental labels was calculated using Pearson’s correlation of pixels within individual segmented cells across all cells in CellProfiler. Similarly, Pearson’s correlation coefficient was used to measure the correlation between features extracted from experimentally and virtually labelled cells across all cells (Feature-PCC). Significance tests were conducted using Welsh’s two-sided t-test (no variance assumptions).

Classification

Cells segmented from the EM dataset (section) were categorised into three classes corresponding to treatment conditions – control, EGF (48hrs), TGF- β1 (48hrs). For a given population, cells were then split into training and test sets using stratified 5-fold cross-validation, maintaining class proportions. Standard scaling was applied to each feature, using the training set to calculate mean and standard deviation parameters. Due to the high dimensionality and redundancies likely present in the data, principal component analysis (PCA) was fit on the training set and performed across the dataset using Minka’s maximum likelihood estimation^43,44 (MLE) to select a number of principal components. Standard scaled and PCA-transformed features then enabled treatment condition classification using a support vector machine (SVM) including a radial basis function kernel. Prediction performance was evaluated using the mean F1-score across the folds.

Populations were defined by the variable marker used for labelling. For each population, we classified using features extracted from (1) only experimentally labelled common markers, (2) experimentally labelled common markers and the variable marker, (3) common markers and a variable marker predicted using ResViT (the top performing model) with DIC input, and (4) experimentally labelled common markers and all variable markers predicted using ResViT, using label-free and common markers input.

Embedding and pseudotime analysis

Unsupervised embeddings were created for the cells in the EM dataset using intensity, radial distribution, and texture features extracted from (1) three common markers; (2) All 11 markers predicted using DIC only; (3) The three common markers experimentally determined, and the 8 variable markers predicted using DIC and the three common markers as input. Features were standard scaled and PCA transformed, using Minka’s maximum likelihood estimation (MLE) to select an optimal number of principal components. Transformed features were then embedded using T-SNE⁴³, UMAP⁴⁴, or PHATE⁴² using default parameters. Extreme outlier cells were removed solely for compact graph visualisation in final figures. Pseudotime analysis was conducted on complete data using scFates⁴⁸ employing a 25-node curve with root determination based on the control treatment condition of the cells.

Code availability

Code used to train and apply models is made available at: https://github.com/CancerSystemsMicroscopyLab/VirtualLabelling

We have also provided sample images and an IPython notebook example to use virtual labelling with other datasets. The same example is provided as a Google Colab notebook which only requires the upload of additional datasets and removes any necessary hardware/dependency set up. Instructions for the use of both notebooks are included in the Github repository.

Code for image preprocessing have also been included. An ImageJ macro is provided to perform background subtraction for DIC images as described in methods. A Python script is provided for the conversion of 16-bit to 8-bit images with contrast enhancement across the whole dataset (per marker as opposed to per image).

Data availability

Raw and pre-processed (as described in methods) images used in Section 1 and Section 2 are provided at: https://zenodo.org/records/xxxxxxx (To be made public upon publication).

Acknowledgements

I.G. and F.V.K are supported by Australian Government Research Training Program (RTP) Scholarships, F.V.K receives a Top-Up award from SPHERE Cancer CAG and Cancer Institute NSW. Y.Z. receives a UNSW TF scholarship. J.G.L. is supported by a University of New South Wales Scientia Research Fellowship, a Ramaciotti Biomedical Research Award, an ARC Development Project grant (DP170103599), NHMRC Ideas Grants (GNT1184009, GNT2012848, GNT2028506), and a Tour de Cure Pioneering Grant (RSP-547-FY2023). This research was undertaken with the assistance of resources and services from the National Computational Infrastructure (NCI), which is supported by the Australian Government.

Author contributions

I.G., E.M., and J.G.L. conceived the paper and project design. Virtual labelling and all downstream analyses conducted by I.G, with input from D.N. Preparation, multiplexed labelling and imaging of DU145 cells conducted by F.V.K and Y.Z. Preparation, labelling, and imaging A549 cells conducted by K.N. Design and writing of the manuscript was led by I.G and J.G.L., with contributions from F.V.K. and significant conceptual development and editing by E.M and F.V.

Chappell, L., Russell, A.J.C., and Voet, T. (2018). Single-cell (multi) omics technologies. Annu Rev Genomics Hum Genet 19, 15–41.
Macaulay, I.C., Ponting, C.P., and Voet, T. (2017). Single-cell multiomics: multiple measurements from single cells. Trends in genetics 33, 155–168.
Butler, A., Hoffman, P., Smibert, P., Papalexi, E., and Satija, R. (2018). Integrating single-cell transcriptomic data across different conditions, technologies, and species. Nat Biotechnol 36, 411–420.
Stuart, T., Butler, A., Hoffman, P., Hafemeister, C., Papalexi, E., Mauck, W.M., Hao, Y., Stoeckius, M., Smibert, P., and Satija, R. (2019). Comprehensive integration of single-cell data. Cell 177, 1888–1902.
Argelaguet, R., Cuomo, A.S.E., Stegle, O., and Marioni, J.C. (2021). Computational principles and challenges in single-cell data integration. Nat Biotechnol 39, 1202–1215.
Lock, J.G., and Strömblad, S. (2010). Systems microscopy: an emerging strategy for the life sciences. Exp Cell Res 316, 1438–1444.
Lichtman, J.W., and Conchello, J.-A. (2005). Fluorescence microscopy. Nat Methods 2, 910–919.
Gut, G., Herrmann, M.D., and Pelkmans, L. (2018). Multiplexed protein maps link subcellular organization to cellular states. Science (1979) 361, eaar7042.
Lin, J.-R., Fallahi-Sichani, M., and Sorger, P.K. (2015). Highly multiplexed imaging of single cells using a high-throughput cyclic immunofluorescence method. Nat Commun 6, 8390.
Seo, J., Sim, Y., Kim, J., Kim, H., Cho, I., Nam, H., Yoon, Y.-G., and Chang, J.-B. (2022). PICASSO allows ultra-multiplexed fluorescence imaging of spatially overlapping proteins without reference spectra measurements. Nat Commun 13, 2475.
Tomimatsu, K., Fujii, T., Bise, R., Hosoda, K., Taniguchi, Y., Ochiai, H., Ohishi, H., Ando, K., Minami, R., Tanaka, K., et al. (2024). Precise immunofluorescence canceling for highly multiplexed imaging to capture specific cell states. Nat Commun 15, 3657. https://doi.org/10.1038/s41467-024-47989-9.
Ounkomol, C., Seshamani, S., Maleckar, M.M., Collman, F., and Johnson, G.R. (2018). Label-free prediction of three-dimensional fluorescence images from transmitted-light microscopy. Nat Methods 15, 917–920.
Christiansen, E.M., Yang, S.J., Ando, D.M., Javaherian, A., Skibinski, G., Lipnick, S., Mount, E., O’Neil, A., Shah, K., Lee, A.K., et al. (2018). In Silico Labeling: Predicting Fluorescent Labels in Unlabeled Images. Cell 173, 792-803.e19. https://doi.org/10.1016/j.cell.2018.03.040.
Helgadottir, S., Midtvedt, B., Pineda, J., Sabirsh, A., B. Adiels, C., Romeo, S., Midtvedt, D., and Volpe, G. (2021). Extracting quantitative biological information from bright-field cell images using deep learning. Biophys Rev 2, 031401.
Cross-Zamirski, J.O., Mouchet, E., Williams, G., Schönlieb, C.-B., Turkki, R., and Wang, Y. (2022). Label-free prediction of cell painting from brightfield images. Sci Rep 12, 1–13.
Cheng, S., Fu, S., Kim, Y.M., Yi, J., and Tian, L. (2020). Deep-learning-enabled virtual immunofluorescence staining based on reflectance microscopy. In 2020 IEEE Photonics Conference (IPC) (IEEE), pp. 1–2.
Sun, G., Liu, S., Shi, C., Liu, X., and Guo, Q. (2023). 3DCNAS: A universal method for predicting the location of fluorescent organelles in living cells in three-dimensional space. Exp Cell Res 433, 113807. https://doi.org/https://doi.org/10.1016/j.yexcr.2023.113807.
Wieslander, H., Gupta, A., Bergman, E., Hallström, E., and Harrison, P.J. (2021). Learning to see colours: Biologically relevant virtual staining for adipocyte cell images. PLoS One 16, e0258546-.
Liu, Y., Yuan, H., Wang, Z., and Ji, S. (2020). Global Pixel Transformers for Virtual Staining of Microscopy Images. IEEE Trans Med Imaging 39, 2256–2266. https://doi.org/10.1109/TMI.2020.2968504.
Guo, S.-M., Yeh, L.-H., Folkesson, J., Ivanov, I.E., Krishnan, A.P., Keefe, M.G., Hashemi, E., Shin, D., Chhun, B.B., Cho, N.H., et al. (2020). Revealing architectural order with quantitative label-free imaging and deep learning. Elife 9, e55502. https://doi.org/10.7554/eLife.55502.
Tonks, S., Hsu, C.-Y., Hood, S., Musso, R., Hopely, C., Titus, S., Krull, A., Doan, M., and Styles, I. (2023). Evaluating virtual staining for high-throughput screening. In 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), pp. 1–5. https://doi.org/10.1109/ISBI53787.2023.10230501.
Ma, J., and Chen, H. (2023). Efficient Supervised Pretraining of Swin-transformer for Virtual Staining of Microscopy Images. IEEE Trans Med Imaging, 1. https://doi.org/10.1109/TMI.2023.3337253.
Lee, G., Oh, J.W., Her, N.G., and Jeong, W.K. (2021). DeepHCS++: Bright-field to fluorescence microscopy image conversion using multi-task learning with adversarial losses for label-free high-content screening. Med Image Anal 70, 101995. https://doi.org/10.1016/j.media.2021.101995.
Weber, L., Lee, B.S., Imboden, S., Hsieh, C.-J., and Lin, N.Y.C. (2023). Phenotyping senescent mesenchymal stromal cells using AI image translation. Curr Res Biotechnol 5, 100120.
Ronneberger, O., Fischer, P., and Brox, T. (2015). U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18 (Springer), pp. 234–241.
Isola, P., Zhu, J.-Y., Zhou, T., and Efros, A.A. (2017). Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1125–1134.
Dalmaz, O., Yurt, M., and Çukur, T. (2022). ResViT: Residual vision transformers for multimodal medical image synthesis. IEEE Trans Med Imaging 41, 2598–2614.
Kramer, B.A., Del Castillo, J.S., Pelkmans, L., and Gut, G. (2023). Iterative Indirect Immunofluorescence Imaging (4i) on Adherent Cells and Tissue Sections. Bio Protoc 13.
Wang, Z., Bovik, A.C., Sheikh, H.R., and Simoncelli, E.P. (2004). Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13, 600–612.
Pearson, K. (1901). LIII. On lines and planes of closest fit to systems of points in space. The London, Edinburgh, and Dublin philosophical magazine and journal of science 2, 559–572.
Carpenter, A.E., Jones, T.R., Lamprecht, M.R., Clarke, C., Kang, I.H., Friman, O., Guertin, D.A., Chang, J.H., Lindquist, R.A., and Moffat, J. (2006). CellProfiler: image analysis software for identifying and quantifying cell phenotypes. Genome Biol 7, 1–11.
Fleischmajer, R. (1967). Epithelial-mesenchymal interactions. Science (1979) 157, 1472–1482.
Hay, E.D. (1995). An overview of epithelio-mesenchymal transformation. Cells Tissues Organs 154, 8–20.
Yang, J., Antin, P., Berx, G., Blanpain, C., Brabletz, T., Bronner, M., Campbell, K., Cano, A., Casanova, J., Christofori, G., et al. (2020). Guidelines and definitions for research on epithelial–mesenchymal transition. Nat Rev Mol Cell Biol 21, 341–352. https://doi.org/10.1038/s41580-020-0237-9.
Xu, Y., and Lu, S. (2015). Transforming growth factor‑β1‑induced epithelial to mesenchymal transition increases mitochondrial content in the A549 non‑small cell lung cancer cell line. Mol Med Rep 11, 417–421.
Fang, C., and Kang, Y. (2021). E-cadherin: context-dependent functions of a quintessential epithelial marker in metastasis. Cancer Res 81, 5800–5802.
Brown, T.C., Sankpal, N. V, and Gillanders, W.E. (2021). Functional implications of the dynamic regulation of EpCAM during epithelial-to-mesenchymal transition. Biomolecules 11, 956.
Loh, C.-Y., Chai, J.Y., Tang, T.F., Wong, W.F., Sethi, G., Shanmugam, M.K., Chong, P.P., and Looi, C.Y. (2019). The E-cadherin and N-cadherin switch in epithelial-to-mesenchymal transition: signaling, therapeutic implications, and challenges. Cells 8, 1118.
Fedorova, O., Parfenyev, S., Daks, A., Shuvalov, O., and Barlev, N.A. (2022). The Role of PTEN in Epithelial–Mesenchymal Transition. Cancers (Basel) 14, 3786.
Usman, S., Waseem, N.H., Nguyen, T.K.N., Mohsin, S., Jamal, A., Teh, M.-T., and Waseem, A. (2021). Vimentin is at the heart of epithelial mesenchymal transition (EMT) mediated metastasis. Cancers (Basel) 13, 4985.
Yan, Y., Zuo, X., and Wei, D. (2015). Concise review: emerging role of CD44 in cancer stem cells: a promising biomarker and therapeutic target. Stem Cells Transl Med 4, 1033–1043.
Moon, K.R., van Dijk, D., Wang, Z., Gigante, S., Burkhardt, D.B., Chen, W.S., Yim, K., Elzen, A. van den, Hirn, M.J., Coifman, R.R., et al. (2019). Visualizing structure and transitions in high-dimensional biological data. Nat Biotechnol 37, 1482–1492. https://doi.org/10.1038/s41587-019-0336-3.
van der Maaten, L., and Hinton, G. (2008). Visualizing data using t-SNE. Journal of machine learning research 9.
McInnes, L., Healy, J., and Melville, J. (2018). Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426.
Miyazono, K. (2009). Transforming growth factor-β signaling in epithelial-mesenchymal transition and progression of cancer. Proceedings of the Japan Academy, Series B 85, 314–323.
Liu, X. (2008). Inflammatory cytokines augments TGF‐β1‐induced epithelial‐mesenchymal transition in A549 cells by up‐regulating TβR‐I. Cell Motil Cytoskeleton 65, 935–944.
Bronte, G., Bravaccini, S., Bronte, E., Burgio, M.A., Rolfo, C., Delmonte, A., and Crinò, L. (2018). Epithelial‐to‐mesenchymal transition in the context of epidermal growth factor receptor inhibition in non‐small‐cell lung cancer. Biological Reviews 93, 1735–1746.
Faure, L., Soldatov, R., Kharchenko, P. V, and Adameyko, I. (2023). scFates: a scalable python package for advanced pseudotime and bifurcation analysis from single-cell data. Bioinformatics 39, btac746.
Loh, C.-Y., Chai, J.Y., Tang, T.F., Wong, W.F., Sethi, G., Shanmugam, M.K., Chong, P.P., and Looi, C.Y. (2019). The E-cadherin and N-cadherin switch in epithelial-to-mesenchymal transition: signaling, therapeutic implications, and challenges. Cells 8, 1118.
Primeaux, M., Gowrikumar, S., and Dhawan, P. (2022). Role of CD44 isoforms in epithelial-mesenchymal plasticity and metastasis. Clin Exp Metastasis 39, 391–406.
Brown, T.C., Sankpal, N. V, and Gillanders, W.E. (2021). Functional implications of the dynamic regulation of EpCAM during epithelial-to-mesenchymal transition. Biomolecules 11, 956.
Lin, J., Fallahi‐Sichani, M., Chen, J., and Sorger, P.K. (2016). Cyclic immunofluorescence (CycIF), a highly multiplexed method for single‐cell imaging. Curr Protoc Chem Biol 8, 251–264.
Nguyen, T., Bui, V., Thai, A., Lam, V., Raub, C.B., Chang, L.-C., and Nehmetallah, G. (2020). Virtual organelle self-coding for fluorescence imaging via adversarial learning. J Biomed Opt 25, 96009.
Jiang, Z., Li, B., Tran, T.N.H.T., Jiang, J., Liu, X., and Ta, D. (2022). Fluo-Fluo translation based on deep learning. Chinese Optics Letters 20, 031701. https://doi.org/10.3788/col202220.031701.
Zheng, A., and Casari, A. (2018). Feature engineering for machine learning: principles and techniques for data scientists (“ O’Reilly Media, Inc.”).
Rubessa, M., and Wheeler, M.B. (2020). Label-free microscopy: A non-invasive new tool to assess gametes and embryo quality. Theriogenology 150, 241–246. https://doi.org/https://doi.org/10.1016/j.theriogenology.2020.01.065.
Gut, G., Herrmann, M.D., and Pelkmans, L. (2018). Multiplexed protein maps link subcellular organization to cellular states. Science (1979) 361, eaar7042.
Gut, G., Herrmann, M.D., and Pelkmans, L. (2018). Multiplexed protein maps link subcellular organization to cellular states. Science (1979) 361, eaar7042.
Kingma, D.P., and Ba, J. (2014). Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.
Carpenter, A.E., Jones, T.R., Lamprecht, M.R., Clarke, C., Kang, I.H., Friman, O., Guertin, D.A., Chang, J.H., Lindquist, R.A., Moffat, J., et al. (2006). CellProfiler: image analysis software for identifying and quantifying cell phenotypes. Genome Biol 7, R100. https://doi.org/10.1186/gb-2006-7-10-r100.
Wang, Z., Bovik, A.C., Sheikh, H.R., and Simoncelli, E.P. (2004). Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13, 600–612.

There is NO Competing Interest.

Supplementary.docx

Download PDF

Version 1

posted

You are reading this latest preprint version

Seeing more with less: Extensible Immunofluorescence (ExIF) accessibly generates high-plexity datasets by integrating standard 4-plex imaging data

Status:

Version 1

Abstract

Figures

Introduction

Results

Single-cell integration through virtual labelling

Applying extensible immunofluorescence to investigate epithelial-mesenchymal cell state plasticity

Classifying cell treatment condition using ExIF integrated data

Mapping EM-state heterogeneity and marker dynamics using ExIF data integration

Discussion

Methods

Cell culture and imaging

Cell culture

EMT activation

Immunofluorescence and Iterative indirect immunofluorescence imaging (4i)

Confocal microscopy

Image analysis, virtual labelling, and downstream quantitative analyses

Pre-processing

Virtual labelling model design and training

Segmentation and Feature extraction

Evaluation

Classification

Embedding and pseudotime analysis

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1