{"paper_id":"f96200ad-7eb6-47ca-aa29-7475265f0771","body_text":"1\nCharacterising cancer-stroma interactions through high-content phenotyping from \nmicroscopy time-lapses \n \nLaura Wiggins1*, Jodie R. Malcolm2, Karen Hogg3, Peter J. O’Toole3, Julie Wilson4, William J. \nBrackenbury2 \n \n1 School of Chemical, Materials and Biological Engineering, University of Sheffield, Sheffield, \nUK \n2 Jack Birch Cancer Research Unit, York Biomedical Research Institute, Department of \nBiology, University of York, Heslington, York, UK \n3 Bioscience Technology Facility, Department of Biology, University of York, Heslington, \nYork, YO10 5DD, UK \n4 Department of Mathematics, University of York, Heslington, York, UK \n \n* To whom correspondence should be addressed. Email: l.wiggins@sheffield.ac.uk \n \nKeywords: Cellular imaging, software, breast cancer \n  \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 3, 2025. ; https://doi.org/10.1101/2025.10.02.680037doi: bioRxiv preprint \n\n 2\nAbstract \n \nUnderstanding how cancer-stromal interactions shape cancer progression requires tools that \ncan capture dynamic phenotypic changes in physiologically relevant conditions. Traditional \napproaches for studying co-culture interactions, such as transcriptomics and flow cytometry, \nprovide valuable insights but are limited by their static nature and reliance on fixed or \ndissociated cells. In contrast, label-free time-lapse microscopy preserves temporal and \nspatial context, enabling observation of live-cell behaviours over time. A major challenge, \nhowever, lies in the analysis of the resulting high-dimensional datasets. Using co-cultures of \nbreast cancer cells and cancer-associated fibroblasts (CAFs) as a model system, we show \nthat the CellPhe toolkit enables label-free identification and phenotypic characterisation of \ndifferent cell types within complex live-cell imaging datasets. Our analysis shows that \nexposure to CAFs drives marked phenotypic shifts in breast cancer cells, including \nelongation, loss of cell–cell adhesion, and redistribution of intracellular components - \nhallmarks of epithelial–mesenchymal transition (EMT). To probe the underlying mechanisms, \nwe performed a Luminex immunoassay on CAF-conditioned media and identified secreted \nanalytes strongly associated with EMT induction. Together, these results highlight how \nautomated phenotyping can be integrated with molecular profiling to identify and \ncharacterise cellular processes shaped by stromal interactions and reveal the signalling \nmediators that drive them. \n \n  \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 3, 2025. ; https://doi.org/10.1101/2025.10.02.680037doi: bioRxiv preprint \n\n 3\nIntroduction \n \nThe tumour microenvironment comprises a diverse array of cell types, including tumour \ncells, immune cells, fibroblasts, endothelial cells, and cancer-associated stromal cells, that \ninteract dynamically to influence tumour progression, metastasis, and therapeutic response\n1. \nWhilst a range of approaches are commonly used to study these interactions, including \ntranscriptomic profiling2, proteomics3, and flow cytometry4, these methods typically provide \nstatic snapshots of dynamic processes and require cell fixation or dissociation, which can \ndisrupt native cellular behaviour. In contrast, live-cell imaging, particularly time-lapse \nmicroscopy, offers direct visualisation of cell co-cultures, capturing real-time cellular \ndynamics and interactions and thus fully preserving both spatial and temporal context. \n One of the primary barriers to using time-lapse microscopy for studying cellular \nbehaviour in co-culture systems is the analytical complexity of the resulting data, particularly \ndue to challenges in tracking and segmenting heterogeneous, interacting cell populations\n5, \nas well as accurate cell type classification to enable cell-type-specific analyses of behaviour \nover time6. A common approach to determine the ground truth identity of cell types in \nmicroscopy images is to label them using fluorescent biomarkers. However, the combination \nof fluorescent staining and frequent light exposure during time-lapse imaging can disrupt \nnormal cell behavior or induce phototoxicity, limiting its applicability in long-term \nexperiments\n7,8. This motivates the use of label-free approaches for time-lapse microscopy, \nsuch as ptychography or phase-contrast microscopy, which enable for extended imaging \nover longer timescales without compromising cell viability or behaviour\n9. As a result, label-\nfree imaging can generate substantially larger data sets, capturing single-cell dynamics over \ntime. Effectively analysing such high-dimensional, time-resolved data requires advanced \ncomputational methods, including single-cell segmentation and tracking, feature extraction, \nand machine learning techniques capable of extracting meaningful biological insights from \ncomplex temporal patterns. \nAs a solution to this, we developed CellPhe10, a toolkit for automated cell \nphenotyping from microscopy timelapses, designed to extract and interpret quantitative \nfeatures from complex cellular behaviours over time. Our previous work has showcased \nCellPhe’s utility in supporting biological research, including the characterisation of \nphenotypic heterogeneity in breast cancer cell responses to chemotherapy10, and multi-class \nphenotypic profiling across a panel of breast cancer cell lines11. In this work, we aim to \ndemonstrate that CellPhe is also a valuable tool for the analysis of cell co-cultures, enabling \ndetailed phenotypic characterisation of cellular interactions. By capturing dynamic \nbehavioural changes at the single-cell level, CellPhe facilitates identification of interaction-\ndriven phenotypes that may otherwise be missed through conventional population-level or \nstatic timepoint analyses. \nAs a proof-of-principle, here we apply the CellPhe toolkit to investigate phenotypic \nchanges in breast cancer cells induced by co-culture with cancer-associated fibroblasts \n(CAFs). The presence of CAFs in the breast cancer tumour microenvironment has been \nwidely reported to correlate with poor clinical prognosis\n12,13, as well as enhanced local \ninvasion and distant dissemination of cancer cells14. Furthermore, CAFs have been shown to \nactively promote epithelial-to-mesenchymal transition (EMT) in breast cancer cells, leading \nto the loss of epithelial characteristics, increased migration and acquisition of mesenchymal \ntraits15. This transition contributes to a more invasive and therapy-resistance phenotype, \ndriving tumour progression and recurrence16. Given that CAFs induce significant phenotypic \nchanges in breast cancer cells, we hypothesised that CellPhe would be capable of \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 3, 2025. ; https://doi.org/10.1101/2025.10.02.680037doi: bioRxiv preprint \n\n 4\nquantifying these changes and facilitating their interpretation in terms of underlying biological \nmechanisms. This capability enables researchers to investigate how intercellular interactions \ninfluence cell state transitions in a high-throughput, label-free manner. By leveraging time-\nresolved features, CellPhe provides a powerful framework for uncovering subtle behavioural \nsignatures that reflect dynamic changes in cellular phenotype, such as those driven by co-\nculture conditions, without the need for molecular labels or endpoint assays. \nAs the phenotypic changes induced in breast cancer cells by CAFs are increasingly \nattributed to complex crosstalk between tumour and stromal cells, mediated by dynamic \nsignalling through secreted cytokines, chemokines, and growth factors, we also performed a \nLuminex immunoassay to profile key signalling molecules within our co-culture experiments. \nThese included TGF/i1β , IL/i1 6, IL/i1 8, VEGF, and HGF, key signalling molecules that are \neither secreted by CAFs or known to play a role in processes such as EMT, cell migration, \nand therapeutic response in breast cancer15,17,18. By correlating these secreted factor profiles \nwith phenotypic features extracted from time-lapse imaging, we present a multi-modal \napproach for investigating the impact of CAF-derived signals on breast cancer cell \nbehaviour, providing insights at both molecular and cellular levels. \n \nMethods \nCell culture. \nThe MDA-MB-231 cells were a gift from M. Djamgoz, Imperial College, London. SkBr3 cells \nwere a gift from J. Rae, University of Michigan and MCF-10A cells were a gift from N. \nMaitland, University of York. MDA-MB-468 and MCF-7 cells were from ATCC. dsRed cancer \nassociated fibroblast (CAF) cells were a gift from V. Speirs, University of Aberdeen. CAFs \nwere stably transduced with recombinant lentivirus and were then sorted using FACS to \nenrich red fluorescent cells, CAFs were also hTERT-immortalised as described previously\n19. \nThe molecular identity of all cell lines was verified by short tandem repeat analysis20. \nAuthenticated cell stocks were stored in liquid nitrogen and thawed for use in experiments. \nThawed cells were subcultured 4-5 times prior to discarding and thawing a new stock to \nensure that the molecular identity of cells was retained throughout. All cell lines, except for \nMCF-10A, were cultured separately in Dulbecco’s Modified Eagle Medium (DMEM) \nsupplemented with 5% fetal bovine serum (FBS) and 4 mM L-glutamine. MCF-10A cells \nwere cultured in DMEM/F12 (Invitrogen) with 5% heat-inactivated horse serum, 0.5 µg/mL \nhydrocortisone, 20 ng/mL human EGF, 10 µg/mL insulin, and 100 ng/mL cholera toxin. To \nminimise imaging artefacts, FBS was filtered using a 0.22 µm syringe filter before use. Cells \nwere incubated at 37°C in plastic filter-cap T-25 flasks and were split at a 1:6 ratio during \npassaging. No antibiotics were added to the cell culture medium. Cells were confirmed to be \nfree of mycoplasma before use in experiments through routine DAPI testing at monthly \nintervals. \n \nCo-culture experiments. \nFor all experiments involving co-culture of MDA-MB-231 or MCF-7 with CAF cells, cell types \nwere monocultured until seeding, where they were seeded together in 24- well plates 24 \nhours prior to imaging. In cases where the concentration of CAFs was increased, cells were \nseeded as cancer cell to CAF ratios of 3:1 (25% CAFs), 1:1 (50% CAFs) and 1:3 (75% \nCAFs), where the number of seeded cells/well remained consistent at 8000. \n \nCAF conditioned medium (CAF-CM). \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 3, 2025. ; https://doi.org/10.1101/2025.10.02.680037doi: bioRxiv preprint \n\n 5\nCAFs were cultured in 25cm² flasks until confluent and then their culture medium (referred to \nas CAF-CM throughout) was collected prior to sub-culturing. MDA-MB-231 or MCF-7 cells \nalready seeded into 24-well plates then received a media change where either the collected \nCAF-CM was added or standard DMEM as a control. Cells were incubated in this media for \n24 hours prior to imaging.  \n \nLuminex experiments. \nSample preparation. \nThe MILLIPLEX® Human Circulating Cancer Biomarker Magnetic Bead Panel 1 \n(HCCBP1MAG-58K) and TGF\nβ 1 Single Plex Magnetic Bead Kit (TGFBMAG-64K-01) were \nused to prepare medium samples for the Luminex immunoassay. It was necessary to seed \nsamples into two separate 96-well plates prior to preparation for the Luminex due to the \nTGFβ 1 single plex requiring an acid pre-treatment that is incompatible with the multi-plex \npanel. Tested samples included conditioned medium collected from MDA-MB-231, MCF-7 \nand CAF cells as well as from MDA-MB-231 and MCF-7 cells that had been cultured in CAF-\nCM for 72 hours prior to sample collection. Experimental variation was handled by testing 5 \nreplicate wells for each sample. Biological variation was handled through replicate \nexperiments, with samples collected from three separate experiments carried out on different \nweeks. Medium samples were centrifuged prior to seeding to ensure complete removal of \ncells or debris that would interfere with assay results. DMEM was added to background, \nstandard and control wells to control for variation in analyte production as a result of sample \nculture medium. Magnetic beads for all analytes were mixed together in a mixing bottle \nprovided within the kit and 25\nμ l of this solution was added to all background, standard, \ncontrol and sample wells. Plates were placed on a plate shaker and left to incubate overnight \nat 4°C. The following day, all wells were washed three times in the provided wash buffer \nfollowing the plate washing instructions in the HCCBP1MAG-58K protocol and 25\nμ l \ndetection antibodies added to each well. Plates were then wrapped in foil and placed on a \nplate shaker at room temperature to incubate for 1 hour. 25μ l of Streptavidin-Phycoerythrin \nwas added to each well and plates re-wrapped in foil and placed on the plate shaker for a \nfurther 30 minute incubation. Well contents was then removed and plates washed a further \nthree times prior to being read on the Luminex 200 system.\n  \n \nData exportation and analysis. \nData exported from the Luminex xPonent software was analysed in R using the drLumi \nand limma packages21,22. Analysis involved use of median fluorescent intensity (MFI) \nvalues and expected concentrations from standard wells for the fitting of standard curves. A \n5-parameter log-logistic function was used for standard curve fitting, or a 4-parameter log-\nlogistic function in cases where 5-parameter fitting did not converge. Models were then used \nfor the prediction of analyte concentrations from raw MFI values for all samples. For any \nsample where concentration was predicted to be lower than the limit of quantification, the \nconcentration for this sample was instead set as the lower limit. \n \nAnalyte gene mapping and EMT cross-referencing. \nGene symbols corresponding to the protein analytes included in the MILLIPLEX® Human \nCirculating Cancer Biomarker Magnetic Bead Panel 1 panel were identified using \nGeneCards: The Human Gene Database (genecards.org). Analytes exhibiting statistically \nsignificant changes in concentration in the conditioned medium of MCF-7 or MDA-MB-231 \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 3, 2025. ; https://doi.org/10.1101/2025.10.02.680037doi: bioRxiv preprint \n\n 6\ncells following treatment with CAF-CM were selected for further analysis. These analytes \nwere cross-referenced against two independent EMT gene datasets: msigdb M5930 \nHallmark Epithelial Mesenchymal Transition (gsea-msigdb.org) and EMTome (emtome.org), \nboth of which contain curated lists of genes associated with EMT processes.  \n  \nImage acquisition and exportation. \nOn the day of imaging, cells were placed onto the Phasefocus Livecyte 2 (Phasefocus \nLimited, Sheffield, UK) to incubate for 30 minutes prior to image acquisition to allow for \ntemperature equilibration. One 500μ m x 500μ m field of view per well was imaged to capture \nas many cells, and therefore data observations, as possible. Selected wells were imaged in \nparallel for 48 hours at 20x magnification with 6 minute intervals between frames, resulting in \nfull time-lapses of 481 frames per imaged well. In cases where fluorescence was used, \nphase and fluorescence images were acquired in parallel for each well. Phasefocus’ Cell \nAnalysis Toolbox® software was used for image processing, including background noise \nreduction with the rolling ball algorithm, as well as cell segmentation and tracking. \n \nData analyses. \nCellPhe analyses. \nAll phenotypic characterisation of cells was performed using the CellPhe toolkit (Python \npackage, version 0.4.2). Within CellPhe, segmentation was performed with the integrated \nCellPose (cyto3 model) implementation, and tracking with the integrated TrackMate \nimplementation (SparseLAP algorithm). Feature extraction was carried out using the \ncell_features and time_series_features functions, and optimal separation \nthresholds were determined using the calculate_separation_scores and \noptimal_separation_features functions. The full list of features extracted through \nCellPhe is provided in Wiggins et al. (2023)\n10. \n \nClassification and clustering. \nCell type classification was carried out using the classify_cells function from the CellPhe \nPython package (v0.4.2), which implements an XGBoost-based supervised classification \nframework. Separate binary classifiers were trained for MDA-MB-231 vs. CAF (training set: n \n= 814 MDA-MB-231, n = 680 CAF) and MCF-7 vs. CAF (training set: n = 1121 MCF-7, n = \n557 CAF). Model performance was evaluated on independent test sets (n = 557 MDA-MB-\n231, n = 255 CAF; n = 557 MDA-MB-231, n = 255 CAF) and reported using accuracy and \nconfusion matrices. Principal component analysis (PCA) was applied to z-scored feature \ndata using scikit-learn. Independent test sets and co-culture datasets were projected into the \nPCA space derived from monoculture training data, enabling comparison of phenotypic \nprofiles and visualisation of population-level shifts. To account for domain shift observed in \nco-culture experiments, unsupervised clustering was performed in PCA-reduced space using \nk-means (k = 2). Cluster identities were assigned based on their proximity to monoculture \ncentroids, and validated against dsRed CAF fluorescence intensity as ground truth. \nStatistical analyses. \nAll statistical analyses were carried out in Python (v3.12.3) using the SciPy, NumPy, and \nstatsmodels libraries unless otherwise stated. For comparisons of phenotypic features \nbetween conditions, nonparametric Kruskal–Wallis (KW) tests were applied, followed by post \nhoc pairwise tests with False Discovery Rate (FDR) correction to account for multiple \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 3, 2025. ; https://doi.org/10.1101/2025.10.02.680037doi: bioRxiv preprint \n\n 7\ncomparisons. To identify monotonic trends in phenotypic features across increasing CAF \ndensities, Spearman’s rank correlation coefficients (ρ ) were computed; features with |ρ | ≥  0.2 \nand p < 0.05 were considered to show consistent directional change. For CAF-CM \nexperiments, two-sample t-tests were used to assess differences in feature distributions \nbetween control and CAF-CM–treated cells, and results were visualised using volcano plots \nof log₂  fold change versus –log₁₀  p-value. For fluorescence intensity comparisons between \nclassification- and clustering-based approaches, mean values ± standard error of the mean \n(SEM) were reported, and differences were interpreted in the context of CAF dsRed signal \nas a ground truth validation measure. In all cases, statistical significance thresholds were set \nat p ≤  0.05 unless otherwise specified.  \nStatistical significance is denoted in figures as:  p > 0.05 (n.s), p ≤  0.05 (*), p ≤  0.01 (**), and \np ≤  0.001 (***), p ≤  0.001 (****). \nResults \nCellPhe is an open-source toolkit for automated single-cell phenotyping from microscopy \ntime-lapse data. Previous studies have demonstrated its utility in supporting biological \nresearch, including the characterisation of phenotypic heterogeneity in breast cancer cell \nresponses to chemotherapy\n10, and multi-class phenotypic profiling across a panel of breast \ncancer cell lines11. The aim of this study was to assess whether CellPhe can extend beyond \nmonoculture applications and enable accurate phenotypic analysis of mixed co-cultures of \ncells, allowing quantification of cell-cell interactions without the need for fluorescent labelling. \nAs such, the experimental design consisted of two co-culture setups, in which CAFs were \ncultured with either MDA-MB-231 or MCF-7 breast cancer cells. ds-Red-labelled CAFs were \nused to provide ground truth information on true cell identity throughout the study, however, \nonly label-free phenotypic features were used for all characterisation and classification \nanalyses. The subsequent results focus on identifying phenotypic changes in MDA-MB-231 \nand MCF-7 cells induced by co-culture with CAFs, with the goal of elucidating potential \ninteraction mechanisms and associated signalling pathways. \nPhenotypic characterisation of MDA-MB-231, MCF-7 and CAFs from microscopy time-\nlapse videos. \nTo assess whether the phenotypic features extracted using CellPhe are sufficient to \ncharacterise breast cancer cell subtypes, time-lapse ptychography images of the breast \ncancer cell lines MDA-MB-231 and MCF-7, as well as CAFs, were acquired for each in \nmonoculture, and time series features were extracted in preparation for supervised \nclassification. Representative time-lapse images at 0h, 24h and 48h are provided in Figure \n1a. CellPhe’s feature selection approach was then applied to reduce the initial set of 1,111 \nphenotypic time-series features to those most discriminative for two classification tasks: \nMDA-MB-231 vs. CAF and MCF-7 vs. CAF. This method calculates a measure of separation \nfor each feature, defined as the ratio of between-groups variance to within-groups variance, \nand applies elbow-based thresholding to identify the most informative features. For the \nMDA-MB-231 vs. CAF comparison, 159 features exceeded the optimal separation threshold \nof 0.4, while 153 features exceeded an optimal threshold of 0.75 for the MCF-7 vs. CAF \ncomparison. The features are categorised into texture, shape, size, and local cell density, \nwith the following feature proportions: For MDA-MB-231 vs. CAF, texture accounted for 77%, \nshape for 10%, density for 9%, and size for 4%. For MCF-7 vs. CAF, texture accounted for \n65%, shape 21%, density 10%, and size 5% (Supplementary Figure 1, Supplementary \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 3, 2025. ; https://doi.org/10.1101/2025.10.02.680037doi: bioRxiv preprint \n\n 8\nTable 1 and 2). The refined list of features was used as input for principal component \nanalysis (PCA), which identified distinct cell populations along PC1 for both MDA-MB-231 \nvs. CAF and MCF-7 vs. CAF. The variance explained by PC1 and PC2 was 52% and 11% \nfor MDA-MB-231 vs. CAF, and 67% and 8% for MCF-7 vs. CAF, respectively (Figure 1b). \nIndependent test sets were projected onto the PCA space to assess whether the identified \ngroupings were preserved in new data (n = 557 for MDA-MB-231, 1121 for MCF-7, and 255 \nfor CAF). The projected points consistently aligned with the correct cell type clusters, \ndemonstrating the generalisability and robustness of these phenotypic features in \ndistinguishing cell populations (Figure 1b). \n CellPhe’s XGBoost-based classification fram ework was used to train models capable \nof distinguishing between MDA-MB-231 and CAF, as well as between MCF-7 and CAF. \nWhen applied to independent test sets, the MDA-MB-231 vs. CAF model achieved \nclassification accuracies of 96% for MDA-MB-231 (537/557 classified as MDA-MB-231) and \n99% for CAF (253/255 classified as CAF), while the MCF-7 vs. CAF model achieved \naccuracies of 99% for MCF-7 (1114/1121 classified as MCF-7) and 100% for CAF (Figure \n1c). \n  \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 3, 2025. ; https://doi.org/10.1101/2025.10.02.680037doi: bioRxiv preprint \n\n \n \n \nFigure 1. CellPhe enables phenotypic characterisation and classification of MDA-MB-231, MCF-7 and \nCAFs. a) Representative ptychographic images of MDA-MB-231, MCF-7 and CAF cells, demonstrating \nobservable differences in their phenotypes. b) PCA scores plots for i. MDA-MB-231 vs. CAFs and ii. MCF-7 vs. \nCAFs. Left panels show training data colour-coded by true class labels, while right panels show independent test \nsets projected into the same PCA space, confirming the preservation of phenotypic separation. c) Confusion \nmatrices for XGBoost classification of i. MDA-MB-231 vs. CAF (537/557 and 253/255 correctly classified) and ii. \nMCF-7 vs. CAF (1114/1121 and 255/255 correctly classified). \n \n  \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 3, 2025. ; https://doi.org/10.1101/2025.10.02.680037doi: bioRxiv preprint \n\n 10\nMDA-MB-231, MCF-7 and CAF classification from time-lapse videos of co-cultures. \nThe successful performance of the MDA-MB-231 vs. CAF and MCF-7 vs. CAF classifiers \nsuggested they would provide a useful tool for facilitating cell type identification from time-\nlapse images of co-cultures. Two sets of co-culture time-lapse experiments were conducted, \none containing 50% MDA-MB-231 and 50% CAFs, and another containing 50% MCF-7 and \n50% CAFs. For co-culture experiments, dsRed-labelled CAFs were used, and ptychography \nand red fluorescence images captured in parallel to provide ground truth labels for model \nvalidation. Representative phase and fluorescence images of the co-cultures are provided in \nFigure 2a, which demonstrate visibility of all cells within the ptychography channel but only \nCAFs in the red fluorescent channel.  \nProjection of co-culture data onto the PCA space from Figure 1b demonstrated a \nloss of clearly defined clusters and a shift in PC1 and PC2 centroids for all cell types, \nindicating a phenotypic shift likely driven by interactions between CAFs and cancer cells \n(Figure 2b). This domain shift suggests that cell behaviour and morphology are altered in \nco-culture conditions compared to monoculture. As a result, classifier performance is \nimpacted, with unclear groupings of cancer cells and CAFs observed following model \ninference (Figure 2b). To address poor classification performance as a result of domain \nshift, k-means clustering was applied to identify two distinct clusters within the PCA space, \nwith each cluster being labelled based on its proximity to the nearest centroid from the \nmonoculture PCA space. As only CAFs express red fluorescence, we used red channel \nintensity to validate cell identities predicted by both classification and clustering approaches. \nThe clustering-based method yielded higher mean red fluorescence in the predicted CAF \ngroup and lower intensities in MDA-MB-231 and MCF-7 cells, indicating improved prediction \nof cell types and fewer misclassifications when using the clustering-based approach (Figure \n2c). \n The intensity for predicted CAFs increased from 4.95×10\n/i1  ± 3.78×10/i1  to 6.01×10/i1  \n± 3.53×10/i1  in the MDA-MB-231 vs. CAF labelling, and from 4.95×10/i1  ± 2.87×10³ to \n5.85×10/i1  ± 3.19×10³ in the MCF-7 vs. CAF labelling. In contrast, fluorescent intensity for \nthe predicted MDA-MB-231 and MCF-7 groups decreased, from 9.99 × 10/i1  ± 7.77 × 10³ to \n5.75 × 10/i1  ± 4.76×10³ for MDA-MB-231 cells, and from 1.79×10/i1  ± 5.15×10² to 1.71×10/i1  ± \n4.50×10² for MCF-7 cells. Examples of cells classified for each cell type are shown in \nSupplementary Figure 2. \n  \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 3, 2025. ; https://doi.org/10.1101/2025.10.02.680037doi: bioRxiv preprint \n\n \nFigure 2. Co-culturing MDA-MB-231 and MCF-7 cells with CAFs induces a phenotypic domain shift. a) \nRepresentative images of MDA-MB-231 and MCF-7 cells co-cultured with dsRed CAFs, where ptychographic \nand fluorescence images were acquired in parallel. b) Co-culture data projected onto the monoculture PCA \nscores plot from Figure 1a for i. MDA-MB-231 and CAF, and ii. MCF-7 and CAF. Left panel shows predicted \nlabels using the monoculture classifier, while the right panel shows labels obtained via k-means clustering and \ncentroid pairing. Points are colour-coded according to predicted class labels. c) Bar plots of mean red fluorescent \nintensity of tracked cells for i. MDA-MB-231 vs. CAF and ii. MCF-7 vs. CAF. Cells are grouped by predicted class \nlabel, with the monoculture classifier used for the plots on the left (labelled “Classification”) and k-means \nclustering with centroid pairing on the right (labelled “Clustering”). The plots demonstrate improved cell type \nidentification using the clustering approach, with increased mean fluorescent intensity for CAFs (4.95×10 /i1  ± \n3.78×10/i1  to 6.01×10/i1  ± 3.53×10/i1  for MDA-MB-231 vs. CAF and 4.95×10/i1  ± 2.87×10³ to 5.85×10/i1  ± 3.19×10³ \nfor MCF-7 vs. CAF) and decreased mean fluorescent intensity for MDA-MB-231 (9.99×10/i1  ± 7.77×10³ to \n5.75×10/i1  ± 4.76×10³) and MCF-7 (1.79×10/i1  ± 5.15×10² to 1.71×10/i1  ± 4.50×10²). Barplots show mean ± SEM. \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 3, 2025. ; https://doi.org/10.1101/2025.10.02.680037doi: bioRxiv preprint \n\n 12\n \n \nCAFs modulate MDA-MB-231 and MCF-7 phenotypes in a density-dependent manner. \nHaving identified a phenotypic domain shift in the co-culture datasets, we next aimed to \npinpoint the specific phenotypic features of MDA-MB-231 and MCF-7 cells that were altered \nby CAFs, by directly comparing their behaviour in monoculture vs. co-culture. MDA-MB-231 \nand MCF-7 cells were co-cultured with increasing densities of CAFs, including 0%, 25% (3:1 \nratio), 50% (1:1 ratio), and 75% (1:3 ratio) CAFs. Representative phase and red \nfluorescence images are provided in Figure 3a. Identification of cells using the k-means and \ncentroid pairing approach showed an expected increase in the proportion of predicted CAFs \nwith higher seeding densities. In co-culture with MDA-MB-231 cells, the proportion of \npredicted CAFs increased from 6% at 0% CAFs to 14%, 40%, and 66% at 25%, 50%, and \n75% CAFs, respectively. Similarly, in co-culture with MCF-7 cells, predicted CAFs rose from \n5% at 0% CAFs to 30%, 56%, and 87% with increasing CAF densities (Supplementary \nFigure 3). Cells identified as MDA-MB-231 and MCF-7 through this approach were then \nused for the subsequent analysis. \nWe applied the Kruskal–Wallis nonparametric test to each CellPhe-derived feature, \nwith False Discovery Rate (FDR) correction to control for multiple comparisons, in order to \nidentify phenotypic changes in breast cancer cells across increasing CAF densities. This \nanalysis revealed 358 significantly altered features in MDA-MB-231 cells and 923 in MCF-7 \ncells, highlighting substantial phenotypic shifts in both lines. To determine the direction of \nthese changes, we next used Spearman’s rank correlation to assess monotonic trends with \nCAF density. Features with an absolute Spearman correlation coefficient greater than 0.2, a \nthreshold chosen to capture modest but consistent responses, were classified as increasing \nor decreasing. In MDA-MB-231 cells, 62 features increased and 135 decreased with rising \nCAF density, while in MCF-7 cells, 148 features increased and 247 decreased. Together, \nthese findings demonstrate that both breast cancer cell lines undergo marked phenotypic \nadaptations in response to co-culture with CAFs (Figure 3b, Supplementary Table 3 and \n4).  \nThe subset of features that passed both the Kruskal–Wallis (KW) significance test \nand the Spearman trend filter were then used as input for PCA. This analysis revealed that \nas CAF density increased, the phenotypes of both MDA-MB-231 and MCF-7 cells \nprogressively diverged from their monoculture states (Figure 3c). The PCA scores plot \nillustrates this shift, with 95% confidence ellipses showing reduced overlap with, and \nincreased traversal away from, the monoculture ellipse as CAF density increases. In both \ncell lines, the primary centroid migration occurred along PC1, which explained 27% of the \nvariance in MDA-MB-231 cells and 50% in MCF-7 cells. This result indicates that CAF \ndensity–driven phenotypic variation is largely captured along the first principal component. \nDirectional feature changes highlighted alterations in cell shape, although different \ndescriptors were more informative for each cell line. The mean length of MDA-MB-231 cells \nincreased from 53.2 ± 14.8 µm in monoculture to 64.6 ± 19.3 µm at 75% CAFs. While the \nKW test indicated significant overall differences (H = 15.77, p = 0.0013), post hoc testing \nshowed that increased elongation was most evident at 50% CAFs (Figure 3d). \n In contrast, for MCF-7 cells, which are typically more rounded and compact, length \nwas less descriptive; instead, changes in sphericity captured their morphological response. \nSphericity decreased significantly in MCF-7 cells with increasing CAF density, from 0.428 ± \n0.067 in monoculture to 0.370 ± 0.088 at 75% CAFs (KW, H = 57.7, p < 0.0001). Post hoc \ntesting confirmed significant reductions at 25% (p = 0.008), 50% (p < 0.0001), and 75% \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 3, 2025. ; https://doi.org/10.1101/2025.10.02.680037doi: bioRxiv preprint \n\n 13\nCAFs (p = 0.0004) compared to monoculture. These results indicate that MCF-7 cells \nprogressively lose their rounded epithelial morphology and adopt a less spherical, more \nmesenchymal-like shape under CAF influence.  \nIn ptychographic imaging, pixel intensity reflects cellular dry mass, allowing texture \nfeatures to capture how intracellular material is organised. One such feature, \nCooc01Var_asc, which measures the temporal ascent of pixel intensity variance, decreased \nsignificantly with increasing CAF density in both breast cancer cell lines. In MDA-MB-231 \ncells, Cooc01Var_asc decreased from  0.348 ± 0.082 in monoculture to 0.278 ± 0.084 at \n75% CAFs (KW, H = 13.327, p = 0.004), while in MCF-7 cells it declined from 0.453 ± 0.116 \nto 0.242 ± 0.054 (KW, H = 198.874, p < 0.0001). This reduction indicates that fluctuations in \ndry mass distribution became less pronounced over time, consistent with a more stable \nintracellular organisation. Such stabilisation is characteristic of a mesenchymal-like state, in \nwhich cells adopt a more polarised and persistent morphology that supports migration. \nFurthermore, local cell density, measured as the number of cells within a defined \nspatial radius around each cell, exhibits a decreasing trend in both MDA-MB-231 (from 0.040 \n± 0.019 in monoculture to 0.018 ± 0.006 at 75% CAFs, KW, H = 47.379, p < 0.0001) and \nMCF-7 cells (from 0.105 ± 0.042 in monoculture to 0.052 ± 0.022 at 75% CAFs, KW, H = \n119.540, p < 0.0001) as the proportion of CAFs increases. This decrease in MDA-MB-231 \ncells can be attributed to their alignment with CAFs (Supplementary Figure 4), which \noccupy a larger area, thereby decreasing the number of neighbouring cells that fall within the \ndefined measurement radius used for density calculation. In the case of MCF-7 cells, this \nreduction in local cell density reflects a loss of their typical epithelial clustering behaviour, \nwith cells dispersing and adopting a more scattered arrangement in response to CAF co-\nculture. \nThese features are hallmarks of a mesenchymal phenotype, characterised by \nelongated morphology, loss of epithelial cell–cell adhesion, reduced local density, and \nstabilisation of intracellular organisation. Together, they indicate a shift toward a polarised \nstate typically associated with migratory capacity. In co-culture, however, these \nmorphological and organisational changes were not paralleled by significant alterations in \nmotility, suggesting that direct CAF contact may constrain increased cancer cell movement \neven as it promotes mesenchymal-like traits. \n  \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 3, 2025. ; https://doi.org/10.1101/2025.10.02.680037doi: bioRxiv preprint \n\n \nFigure 3. Co-culture with CAFs induces concentration-dependent phenotypic changes in MDA-MB-231 \nand MCF-7 cells. a) Representative ptychographic and red fluorescence images for dsRed CAFs co-cultured \nwith i. MDA-MB-231 and ii. MCF-7. Note that 0%, 25%, 50% and 75% CAFs refer to cancer cell to CAF seeding \nratios of 1:0, 3:1, 1:1 and 1:3 respectively. b) Area plots show normalised mean values of features that exhibited \ndirectional trends (absolute Spearman’s ρ  ≥  0.2, p/i1 </i1 0.05) across increasing CAF concentrations (0%, 25%, \n50%, 75%). Features were separated into those that i. increased or ii. decreased with CAF concentration. Each \ncoloured area represents one feature, and the bold black line indicates the average trend across all positively or \nnegatively correlated features. c) PCA score plots for i. MDA-MB-231 and ii. MCF-7 based on features shown to \nsignificantly vary with increasing CAF density, as determined by Kruskal-Wallis and Spearman’s rank correlation. \nPCA scores plots show phenotypic divergence of cancer cells with increasing CAF concentration (0%, 25%, 50%, \n75%). 95% confidence ellipses are shown for each concentration, with the black centroid path highlighting \nprogressive separation from the control (0%) as CAF concentration increases, indicating dose-dependent shifts in \ncancer cell phenotype. \nd) Boxplots of example features that showed positive or negative correlations with \nincreasing CAF density for (i) MDA-MB-231 and (ii) MCF-7 cells. Shown are an increase in mean length for MDA-\nMB-231 cells and a decrease in sphericity for MCF-7 cells, as well as decrease in texture variance and local cell \ndensity for both cell types. \n \n  \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 3, 2025. ; https://doi.org/10.1101/2025.10.02.680037doi: bioRxiv preprint \n\n 15\nCulturing with CAF-conditioned medium induces changes to MDA-MB-231 and MCF-7 \nphenotypes. \nTo assess whether phenotypic changes in cancer cells could be driven by CAF-secreted or -\ndepleted factors, such as chemokines and cytokines, rather than by direct cell-cell contact, \nMDA-MB-231 and MCF-7 cells were cultured in CAF-conditioned medium (CAF-CM), and \ncompared to MDA-MB-231 and MCF-7 cells grown in standard culture medium. \nRepresentative time-lapse images are shown in Figure 4a. Based on visual inspection, \nMDA-MB-231 cells cultured in CAF-CM appeared more elongated, while MCF-7 cells \ndisplayed greater variation in shape, including elongated morphologies uncharacteristic of \nthis line and more typical of mesenchymal-like cells. In addition, MCF-7 cells formed fewer \ncolonies, with individual cells more often migrating away from one another. To quantify both \nthe magnitude and statistical significance of CAF-CM–induced phenotypic changes, log\n₂  fold \nchanges and p-values from two-sample t-tests were calculated for all features extracted \nusing CellPhe. These results are visualised using volcano plots in Figure 4b. The analysis \nshowed that CAF-CM significantly altered the phenotypic profiles of both cell lines. For MDA-\nMB-231 cells, 787 out of 1,111 features (≈  71%) exhibited significant changes (p < 0.01) \nupon culture with CAF-CM, while for MCF-7 cells, 566 out of 1,111 features (≈  51%) were \nsignificantly altered. \nFeatures that underwent significant changes following culture with CAF-CM were \nindicative of a shift toward more mesenchymal and aggressive phenotypes, consistent with \nthose observed during direct co-culture with CAFs (Figure 4c, Supplementary Table 5). \nNotably, both MDA-MB-231 and MCF-7 cells exhibited increased elongation and reduced \nsphericity following exposure to CAF-CM. For example, the mean cell length increased from \n56±14μ m to 65±15μ m in MDA-MB-231 cells (p ≤  0.001), and from 55±8/i1μ m to 57±12/i1μ m in \nMCF-7 cells (p ≤  0.001).  \nIn co-culture, this stabilisation was reflected by reduced variance in pixel intensity \ndistribution over time, whereas in CAF-CM it was captured through a reduction in \nfluctuations of overall cellular mass, with the total ascent in mean pixel intensity across the \ntime series decreasing from 1.80 ± 0.65 to 1.40 ± 0.66 in MDA-MB-231 cells (p ≤  0.001), and \nfrom 1.76 ± 0.55 to 1.38 ± 0.66 in MCF-7 cells (p ≤  0.001). A reduction in fluctuations of \naverage cellular mass over time indicates that cells are no longer continually reorganising \ntheir internal contents, but instead maintain a more fixed intracellular architecture. Such \nstabilisation is a recognised feature of EMT, where cells transition to a polarised, elongated \nstate that supports persistent migration\n23. \nUnlike in co-culture, where movement features remained unchanged, cancer cells \ngrown in CAF-CM exhibited significantly greater migration. This was quantified using the \ntrajectory area, which measures the two-dimensional space a cell covers during the \ntimelapse, with larger trajectory areas indicating that cells travelled further from their starting \nposition, consistent with increased motility. In MDA-MB-231 cells, the mean trajectory area \nincreased from 91 ± 76 \nμ m²  in monoculture to 104 ± 91 μ m²  in CAF-CM (p < 0.01). \nSimilarly, in MCF-7 cells, the trajectory area increased from 11 ± 9 μ m²  to 13 ± 1 μ m² 1 (p < \n0.001). The absence of a corresponding effect in co-culture may reflect the physical \nconstraints imposed by CAFs themselves, which can hinder cancer cell displacement \ndespite promoting mesenchymal-like morphology and intracellular stabilisation. \nDecrease in local cell density for MCF-7 cells cultured in CAF-CM (from 0.23 ± 0.09 \nto 0.19 ± 0.08, p ≤/i1  0.001), as was observed in co-culture data sets, again suggests a loss \nof polarity and the transition towards a more mesenchymal-like state associated with \ndecreased cell-cell adhesion or altered migration patterns. In contrast, MDA-MB-231 cells \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 3, 2025. ; https://doi.org/10.1101/2025.10.02.680037doi: bioRxiv preprint \n\n 16\nexperienced an increase in local cell density (0.068 ± 0.03px⁻ ¹ to 0.079 ± 0.04px⁻ ¹, p ≤/i1  \n0.001). This observation is consistent with the timelapse videos of MDA-MB-231 cells \ncultured in CAF-CM, which showed a tendency for closely positioned cells to align and \nmigrate collectively (Supplementary Figure 5). These findings suggest that CAF-CM not \nonly promotes mesenchymal traits but may also facilitate coordinated movement and local \nclustering of MDA-MB-231 cells. \n  \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 3, 2025. ; https://doi.org/10.1101/2025.10.02.680037doi: bioRxiv preprint \n\n \n \n \nFigure 4. Co-culture with CAF-CM induces phenotypic changes in MDA-MB-231 and MCF-7 cells a) \nRepresentative time-lapse images of MDA-MB-231 and MCF-7 cells cultured in standard culture medium or CAF-\nCM. b) Volcano plots depicting phenotypic changes induced by CAF-CM in i. MDA-MB-231 and ii. MCF-7 cells. \nPoints are coloured blue for significant changes (p ≤  0.01), and grey for non-significant changes. The dashed \nhorizontal line at -log₁₀  p-value = 2 indicates a threshold for statistical significance, while the dashed vertical lines \nat log₂  fold change of ±1 denote a 2-fold change. c) Violin plots showing selected phenotypic features with \nsignificant changes induced by CAF-CM in i. MDA-MB-231 and ii. MCF-7 cells. Significant differences between \ncontrol and CAF-CM conditions are indicated, with statistical significance determined by two-sample t-tests (** =  \np \n≤/i1  0.01 , *** =  p ≤/i1  0.001).  \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 3, 2025. ; https://doi.org/10.1101/2025.10.02.680037doi: bioRxiv preprint \n\n 18\nMicroenvironmental changes in MDA-MB-231 and MCF-7 cells cultured with CAF-CM. \nCAF-CM-induced phenotypic changes in MDA-MB-231 and MCF-7 cells, consistent with an \nincrease in mesenchymal-like characteristics, led us to investigate whether secreted factors \nwithin CAF-CM are associated with EMT. A Luminex multiplex assay of 25 analytes was run \nto determine the presence or absence of certain soluble factors within cultured medium \nsamples. The test panel consisted of magnetic beads for the analysis of AFP, Total PSA, \nCA15-3, CA19-9, MIF, TRAIL, Leptin, IL-6, sFasL, CEA, CA125, IL-8, HGF, sFas, TNF\nα , \nProlactin, SCF, CYFRA 21-1, OPN, FGF2, β -HCG, HE4, TGFα , VEGF-A and TGF-β . \nThe heatmap of scaled MFI values (Figure 5a) showed the presence of two main \nanalyte clusters within the dataset: one for analytes in high abundance within CAF-CM, and \none for analytes with lower abundance. The high-abundance cluster included TGF-β , IL-8, \nOPN, sFas, IL-6, HGF, and VEGF-A. Across both MDA-MB-231 and MCF-7 cells, these \nfactors were present at relatively low levels in controls but increased markedly in CAF-CM. \nFor example, IL-6 rose from 330 ± 257 to 16,453 ± 4565 MFI in MDA-MB-231 cells (log/i1 FC \n= 5.64, q = 0.059) and from 414 ± 622 to 14,739 ± 5999 MFI in MCF-7 cells (log/i1 FC = 5.16, \nq = 0.095). Similarly, IL-8 increased from 3748 ± 2136 to 16,217 ± 878 MFI in MDA-MB-231 \ncells (log/i1 FC = 2.11, q = 0.015) and from 1298 ± 1692 to 14,467 ± 451 MFI in MCF-7 cells \n(log/i1 FC = 3.48, q = 0.013). TGF-β  also showed consistent increases, from 1716 ± 669 to \n6172 ± 706 MFI in MDA-MB-231 cells (log/i1 FC = 1.85, q = 0.010) and from 1760 ± 840 to \n5775 ± 555 MFI in MCF-7 cells (log/i1 FC = 1.71, q = 0.013). Other analytes in this cluster, \nsuch as OPN, sFas, and HGF, showed large fold increases (e.g., OPN increased ~36-fold in \nMDA-MB-231 and ~13-fold in MCF-7), but owing to greater variability between replicates \ntheir changes did not consistently reach significance after FDR correction. VEGF-A exhibited \nonly modest increases in both cell lines (MDA-MB-231: 2328 ± 3019 to 4364 ± 1862 MFI; \nMCF-7: 603 ± 408 to 1736 ± 732 MFI), which were also not significant after correction. \nIn contrast, the second cluster comprised analytes abundant in the medium of MCF-7 \ncells, including TGF\nα , MIF, HE4, FGF2, SCF, Total PSA, β -HCG, Leptin, sFasL, CA125, \nTNFα , TRAIL, AFP, Prolactin, CYFRA 21-1, CA15-3, and CEA. Strikingly, the addition of \nCAF-CM led to reductions in most of these factors, particularly in MCF-7 cells. For example, \nHE4 decreased from 31 ± 3.8 to 14.6 ± 2.5 MFI (log₂ FC = –1.09, q = 0.072), FGF2 from 12.0 \n± 1.1 to 8.9 ± 0.1 MFI (log₂ FC = –0.43, q = 0.093), Leptin from 84.3 ± 25.3 to 19.6 ± 8.8 MFI \n(log₂ FC = –2.10, q = 0.093), TRAIL from 245.9 ± 62.3 to 60.7 ± 32.2 MFI (log₂ FC = –2.02, q \n= 0.088), and AFP from 119.8 ± 24.3 to 46.0 ± 11.0 MFI (log₂ FC = –1.38, q = 0.088). \nProlactin was also reduced, from 165 ± 45.6 to 66.6 ± 27.9 MFI (log₂ FC = –1.31, q = 0.093). \nSeveral analytes, including MIF, CA125, CYFRA 21-1, and CEA, showed downward trends \nbut did not reach significance, with CEA in particular showing little change (151 ± 65.5 to 154 \n± 79.9 MFI, log\n₂ FC ≈  0, q ≈  0.97). In MDA-MB-231 cells, changes were less pronounced, \nalthough SCF increased (10.1 ± 1.9 to 21.2 ± 3.7 MFI, log₂ FC = 1.07, q = 0.14) while sFasL \ndecreased (11.1 ± 1.0 to 8.3 ± 0.8 MFI, log₂ FC = –0.42, q = 0.14). \nBarplots are provided in Figure 5b and Figure 5c to further aid visualisation of the \nchanges to MFI for certain analytes induced by the addition of CAF-CM. Additionally, \nbarplots of EMT-specific analytes identified by cross-referencing soluble factors showing \nsignificant changes in the culture medium after CAF-CM treatment to two independent EMT \nsignature databases (msigdb and EMTome) are shown in Supplementary Figure 6. MFI \nvalues for IL-6, IL-8 and TGF-\nβ  are shown in Figure 5b. These plots show the presence of \nIL-6, IL-8 and TGF-β  within CAF-CM, as well as a highly significant increase within MDA-\nMB-231 and MCF-7 samples where CAF-CM had been added, across all 3 experimental \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 3, 2025. ; https://doi.org/10.1101/2025.10.02.680037doi: bioRxiv preprint \n\n 19\nreplicates. These plots suggest an additive effect in which addition of CAF-CM to MDA-MB-\n231 and MCF-7 cells results in a culture medium that contains the summation of secretions \nfrom each cell type. MFI values for TGF\nα , leptin and CA-125 are shown in Figure 5c and \nSupplementary Figure 6. The secretion of leptin and CA-125 from MCF-7 cells was \nsignificantly decreased following culture with CAF-CM, compared to MCF-7 cells cultured in \nnormal medium for the same period, suggesting an inhibitory effect on such analytes \nmediated by factors secreted by CAFs. The same was observed for TGF\nα , with significant \ndecrease in production by MCF-7 cells following culture with CAF-CM. Interestingly, no \nsignificant decrease in TGFα  was observed for MDA-MB-231 cells following culture with \nCAF-CM for experimental replicates 1 and 2, although a significant decrease was observed \nin replicate 3 where the original MFI for TGF\nα  within the control MDA-MB-231 sample was \nmuch higher than for the other replicates. This discrepancy could suggest that inhibition of \nTGFα  in MDA-MB-231 cells by CAFs is dependent on the concentration of TGFα  produced \nby the cancer cells, with greater concentrations resulting in a greater inhibitory response by \nCAFs. Whereas, the observed decrease in EMT-specific analytes in the medium of MCF-7 \ncells following CAF-CM treatment may suggest a context-dependent modulation of epithelial \nplasticity, and may reflect a more nuanced cell-specific response to stromal signalling.  \n \nIn summary, the results of this study have shown how CellPhe’s phenotypic profiling is able \nto not only accurately distinguish breast cancer cell populations, but also reveal important \nCAF-driven, density-dependent EMT-like transitions, marked by elongation, intracellular \nreorganisation and increased migration, captured through time-lapse imaging. These shifts, \nwhich were replicated by CAF-CM and validated by Luminex experiments, underscore the \nsecretory influence of CAFs in modulating cancer cell behavior and affirm CellPhe’s power in \ndecoding dynamic phenomic signatures of malignancy.   \n  \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 3, 2025. ; https://doi.org/10.1101/2025.10.02.680037doi: bioRxiv preprint \n\n \n \nFigure 5. CAF-CM induces changes in analyte production in MDA-MB-231 and MCF-7 cells. a) Heatmap to \naid visualisation of differences in analyte production of MDA-MB-231 and MCF-7 induced by culture with CAF-\nCM. Note that MFI values were first averaged to provide mean values for each of three experimental replicates, \nvalues were then normalised to allow changes on different scales to be compared. The heatmap demonstrates \nthat addition of CAF-CM increases the presence of certain analytes within MDA-MB-231 and MCF-7 culture \nmedium, such as TGF-B, IL-6 and OPN. The addition of CAF-CM reduces the presence of certain analytes within \nMCF-7 medium such as CA15-3, TGF-A and TNF-A. \nb) The addition of CAF-CM significantly increased the \npresence of IL-6, IL-8 and TGF-B in MDA-MB-231 and MCF-7 medium samples across all three experimental \nreplicates, highlighting changes to cancer cell microenvironment induced by CAFs. c) The addition of CAF-CM \nsignificantly decreased the presence of TGF-A, leptin and CA-125 in MCF-7 cells across all three experimental \nreplicates, suggesting a cross-talk between cancer cells and CAFs in which secretions from CAFs inhibit certain \nsecretions from cancer cells. \n  \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 3, 2025. ; https://doi.org/10.1101/2025.10.02.680037doi: bioRxiv preprint \n\n 21\nDiscussion \nUsing microscopy to assess changes to cell behaviour induced by co-culturing with different \ncell types is currently a challenging task due to the complexities that arise when trying to \nseparate out cell types following image acquisition. The results presented here show that our \ntoolkit for automated cell phenotyping, CellPhe, is able to accurately classify cell types within \nco-cultures using their phenotypic characteristics in order to isolate populations for \ndownstream analysis. Current approaches for isolation of specific cell populations include \nflow cytometry and fluorescence-activated cell sorting, though such approaches involve \nconsideration of only a handful of pre-determined metrics for characterisation of cell types \nand have the potential to induce stress within cells\n24. In contrast, CellPhe facilitates accurate \nisolation of cell types using a rich panel of label-free features, enabling interpretable feature \nselection without negatively impacting cell health. \nCulturing MDA-MB-231 and MCF-7 cells with CAFs as well as CAF-CM induced \nphenotypic changes in both cell lines, with cells displaying more elongated cell shape, \ndecreased colony formation, and stabilisation of intracellular components. Together, these \nhallmarks are consistent with a shift towards a more aggressive, mesenchymal-like \nphenotype. These findings strongly suggest that CAFs drive EMT in both cancer cell lines, in \nline with previous reports of CAF-induced EMT in breast cancer\n15. Our findings reinforce the \nemerging view that CAFs act as central orchestrators of epithelial plasticity within the tumour \nmicroenvironment, shaping breast cancer cell behaviour through EMT induction and thereby \npromoting aggressive disease phenotypes. Importantly, we also demonstrate a new \nmicroscopy-based approach for identifying these dynamic changes directly from time-lapse \ndata. \nOur findings showed that the magnitude of these phenotypic shifts was modulated by \nCAF abundance, with higher CAF density driving progressively larger deviations from \nmonoculture phenotypes. This observation aligns with previous findings that CAF-rich \ntumours are often more aggressive \n25–27, and provides a mechanistic basis for how CAF \ndensity in the tumour microenvironment may scale EMT induction. Interestingly, while cells \nexhibited mesenchymal-like morphologies in direct co-culture, motility was not increased \nunder these conditions, in contrast to CAF-CM where motility increased significantly. This \ndifference suggests that CAF contact imposes physical constraints even as it promotes \nEMT-like states, whereas soluble signalling alone promotes cell dispersal and migration. \nThe Luminex immunoassay provided additional, complementary information to the \nphenotypic metrics extracted from time-lapse images, providing a list of biomarkers that may \nmediate the phenotypic changes induced in cancer cells by CAFs. The use of a pre-existing \nmultiplex panel meant that biomarkers with limited associations to CAFs within the literature, \nsuch as CA-125, were also able to be assessed with no prior hypothesis of their involvement \nin CAF-cancer cell interactions. The Luminex assay showed the presence of analytes known \nto be secreted by CAFs within CAF-CM such as TGF-B, IL-6 and IL-8, with all of these \nanalytes also present within medium collected following culturing of MDA-MB-231 and MCF-\n7 cells in CAF-CM. Such analytes were at relatively low concentrations within the control \nmedium samples, demonstrating how the presence of CAF secretions alters the \nmicroenvironment of cancer cells and suggesting their potential role in inducing the observed \nphenotypic changes in cancer cells following culture with CAFs.  \nBy cross-referencing analytes significantly altered by CAF-CM treatment with curated \nEMT signature databases (MSigDB and EMTome), we identified a subset of factors with \nstrong associations to EMT regulation. Notably, EMT-promoting analytes such as TGF-B, IL-\n6, and IL-8 were consistently elevated in CAF-conditioned medium and in the medium \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 3, 2025. ; https://doi.org/10.1101/2025.10.02.680037doi: bioRxiv preprint \n\n 22\ncollected from MDA-MB-231 and MCF-7 cells cultured with CAF-CM, while several analytes \ncharacteristic of epithelial states, including those secreted by MCF-7 cells, were reduced \nfollowing CAF-CM treatment. This dual pattern, upregulation of EMT-inducing signals \ntogether with suppression of epithelial-associated factors, provides evidence that CAF-\nderived secretions remodel the tumour microenvironment in a way that promotes epithelial \nplasticity and drives cancer cells towards mesenchymal-like phenotypes. \nIn summary, this study establishes CellPhe as a framework for analysing phenotypic \nchanges in cancer–stromal interactions directly from label-free time-lapse imaging. Through \nautomated feature extraction, classification, and clustering, CellPhe enabled identification of \nbreast cancer cells within co-cultures, allowing their phenotypes to be directly compared with \nmonoculture breast cancer cells and revealing hallmarks of EMT driven by both contact and \nsoluble CAF signals. The toolkit generalises across cell types and extends to complex \nmulticellular systems, providing a scalable means to study tumour–microenvironment \ninteractions. Coupling imaging-based phenotyping with immunoassays further linked \nphenotypes to CAF-derived EMT factors, demonstrating how CellPhe can generate \nmechanistic hypotheses for molecular validation. \nCompeting interests \nThe authors declare that they have no competing interests. \n \nAcknowledgements \nThe authors would like to acknowledge Prof Valerie Speirs, who kindly shared the CAF cell \nline used in this study. The authors also gratefully acknowledge assistance from the Imaging \nand Cytometry Lab in the Bioscience Technology Facility at the University of York. This work \nwas supported by the MRC (MR/X018067/1), BBSRC (BB/Y513970/1), and the Wellcome \nTrust (310891/Z/24/Z). \n \nAuthors' contributions \nLW, JM, KH, PO’T, JW and WB contributed to the conception and design of the work. LW, \nJM, KH, PO’T, JW and WB contributed to acquisition, analysis, and interpretation of data for \nthe work. LW, JM and WB contributed to drafting the work and revising it critically for \nimportant intellectual content. All authors approved the final version of the manuscript. \n \nReferences \n \n1. Soysal, S. D., Tzankov, A. & Muenst, S. E. Role of the Tumor Microenvironment in \nBreast Cancer. Pathobiology 82, 142–152 (2015). \n2. Gonzalez, E. A. A. Deciphering Patterns of Cell-Cell Interactions and Communication in \nMulticellular Organizations. (2023). \n3. Wang, T. et al. scProAtlas: an atlas of multiplexed single-cell spatial proteomics imaging \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 3, 2025. ; https://doi.org/10.1101/2025.10.02.680037doi: bioRxiv preprint \n\n 23\nin human tissues. Nucleic Acids Res. 53, D582–D594 (2025). \n4. Carretta, M. et al. Dissecting tumor microenvironment heterogeneity in syngeneic \nmouse models: insights on cancer-associated fibroblast phenotypes shaped by \ninfiltrating T cells. Front Immunol 14, 1320614 (2023). \n5. Meijering, E., Dzyubachyk, O. & Smal, I. Methods for cell and particle tracking. Methods \nEnzymol 504, 183–200 (2012). \n6. Sommer, C. & Gerlich, D. W. Machine learning in cell biology - teaching computers to \nrecognize phenotypes. J Cell Sci 126, 5529–5539 (2013). \n7. Laissue, P. P., Alghamdi, R. A., Tomancak, P., Reynaud, E. G. & Shroff, H. Assessing \nphototoxicity in live fluorescence imaging. Nat Methods 14, 657–661 (2017). \n8. Stephens, D. J. & Allan, V. J. Light mi croscopy techniques for live cell imaging. Science \n300, 82–86 (2003). \n9. Suman, R. et al. Label-free imaging to study phenotypic behavioural traits of cells in \ncomplex co-cultures. Sci Rep 6, 22032 (2016). \n10. Wiggins, L. et al. The CellPhe toolkit for cell phenotyping using time-lapse imaging and \npattern recognition. Nat Commun 14, 1854 (2023). \n11. Wiggins, L. et al. CellPhePy: A python implementation of the CellPhe toolkit for \nautomated cell phenotyping from microscopy time-lapse videos. J Microsc (2025) \ndoi:10.1111/jmi.13416. \n12. Kim, H. S. et al. Cancer-Associated Fibroblasts Together with a Decline in CD8+ T Cells \nPredict a Worse Prognosis for Breast Cancer Patients. Ann Surg Oncol 31, 2114–2126 \n(2024). \n13. Shah, S. H. The Contribution of Cancer Associated Fibroblasts in Breast Cancer \nProgression. (2016). \n14. Li, Y., Wang, C., Huang, T., Yu, X. & Tian, B. The role of cancer-associated fibroblasts \nin breast cancer metastasis. Front Oncol 13, 1194835 (2023). \n15. Yu, Y. et al. Cancer-associated fibroblasts induce epithelial-mesenchymal transition of \nbreast cancer cells through paracrine TGF-β  signalling. Br J Cancer 110, 724–732 \n(2014). \n16. Yuan, J. et al. Decoding tumor microenvironment: EMT modulation in breast cancer \nmetastasis and therapeutic resistance, and implications of novel immune checkpoint \nblockers. Biomed Pharmacother 181, 117714 (2024). \n17. Sun, X. et al. IL-6 secreted by cancer-associated fibroblasts induces tamoxifen \nresistance in luminal breast cancer. Oncogene 33, 4450 (2014). \n18. Schneider, B. P. & Sledge, G. W., Jr. Drug insight: VEGF as a therapeutic target for \nbreast cancer. Nat Clin Pract Oncol 4, 181–189 (2007). \n19. Nash, C. E. et al. Development and characterisation of a 3D multi-cellular in vitro model \nof normal human breast: a tool for cancer initiation studies. Oncotarget 6, 13731–13741 \n(2015). \n20. Masters, J. R. et al. Short tandem repeat profiling provides an international reference \nstandard for human cell lines. Proc Natl Acad Sci U S A 98, 8012–8017 (2001). \n21. Sanz, H. et al. drLumi: An open-source package to manage data, calibrate, and conduct \nquality control of multiplex bead-based immunoassays data analysis. PLoS One 12, \ne0187901 (2017). \n22. Ritchie, M. E. et al. limma powers differential expression analyses for RNA-sequencing \nand microarray studies. Nucleic Acids Res. 43, e47 (2015). \n23. Datta, A. et al. Cytoskeletal Dynamics in Epithelial-Mesenchymal Transition: Insights \ninto Therapeutic Targets for Cancer Metastasis. Cancers (Basel) 13, (2021). \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 3, 2025. ; https://doi.org/10.1101/2025.10.02.680037doi: bioRxiv preprint \n\n 24\n24. Pfister, G., Toor, S. M., Sasidharan Nair, V. & Elkord, E. An evaluation of sorter induced \ncell stress (SICS) on peripheral blood mononuclear cells (PBMCs) after different sort \nconditions - Are your sorted cells getting SICS? J Immunol Methods 487, 112902 \n(2020). \n25. Cords, L. et al. Cancer-associated fibroblast phenotypes are associated with patient \noutcome in non-small cell lung cancer. Cancer Cell 42, 396–412.e5 (2024). \n26. Min, K.-W. et al. Cancer-associated fibroblasts are associated with poor prognosis in \nsolid type of lung adenocarcinoma in a machine learning analysis. Sci. Rep. 11, 16779 \n(2021). \n27. Wang, X. et al. Overcoming cancer treatment resistance: Unraveling the role of cancer-\nassociated fibroblasts. J Natl Cancer Cent 5, 237–251 (2025). \n \n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprintthis version posted October 3, 2025. ; https://doi.org/10.1101/2025.10.02.680037doi: bioRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}