Microscopy-based cancer cell classification traditionally relies on cell-based morphological features, while subcellular organelle organization remains underutilized. Existing machine learning methods often require manual preprocessing and handcrafted feature extraction, limiting scalability and introducing user bias. This study proposes an automated, interpretable, and organelle-focused deep learning framework for classifying breast cancer cell lines from high-resolution fluorescence microscopy images. We developed an end-to-end framework that incorporates patch-based sampling, sparsity filtering, and a channel-wise intermediate fusion strategy to independently extract and integrate organelle-specific features. Model interpretability was assessed using Grad-CAM visualizations and single-organelle classifier analyses. The framework was evaluated on fluorescence microscopy images from six breast cancer cell lines using 5-fold cross-validation. The proposed framework achieved a classification accuracy of 97.1 ± 1.1

During metastatic progression, breast cancer cells exhibit behaviors such as organ-specific tropism and varied responses to metastatic inhibitors. However, the similarity in cellular morphology between metastatic subtypes limits the effectiveness of conventional microscopy for their identification. While fluorescence microscopy combined with deep learning enables automated identification of nonvisible features, such models require large annotated datasets. Therefore, limited annotations and inconsistent data handling commonly hinders network generalizability and task performance. To address this, we utilize self-supervised learning (SSL) with sparsity-filtered patch-based preprocessing for user-agnostic classification of metastatic tropism and drug-treated triple-negative MDA-MB-231 variants. A threshold-based pipeline first extracts non-sparse 2D patches from 3D multi-channel confocal images for mitochondria and early endosomes. SSL is then applied using two pretext tasks: a jigsaw task that learns spatial structure by predicting permutations of shuffled sub-patches, and contrastive learning task that learns augmentation invariance by distinguishing similarity between augmented views of the same patch and dissimilarity between unrelated patches. The learned representation is then fine-tuned for two downstream classification tasks: metastatic tropism subtype identification, including parental, lung-tropic, and bone-tropic cells; and characterization of cellular response to varying concentrations of the metastatic inhibitor, MGH-CP1. From these results, we demonstrate that SSL-pretrained models improve classification of metastatic subtypes based on mitochondrial and endosomal distribution, outperforming gold-standard pretraining approaches. This workflow provides a scalable approach for microscopic cancer image analysis and may support future efforts in automated cancer diagnostics. Furthermore, the SSL framework enables generalization to specialized microscopic analysis tasks without requiring labeled datasets, thus making it suitable for broader applications.

Supplemental Table 2. Primer sequences for genotyping, qualitative reverse transcriptase-PCR, and oligonucleotides for cloning.

Background/Objectives: Epithelioid Hemangioendothelioma (EHE) is an ultra-rare, metastatic vascular sarcoma with limited therapeutic options. The hallmark of EHE is a chromosomal translocation that produces the WWTR1-CAMTA1 gene fusion, encoding the aberrant transcriptional regulator TAZ-CAMTA1. Given its central role in the EHE initiation and progression, TAZ-CAMTA1 represents a compelling therapeutic target. Methods and Results: In this study, we identified AMP-activated protein kinase (AMPK) as one of several proteins capable of repressing the TAZ-CAMTA1 transcriptional activity in NIH3T3 and HEK293 cell lines. The pharmacologic activation of AMPK inhibited the proliferation of EHE cell lines without inducing apoptosis; however, in contrast to the NIH3T3 and HEK293 models, AMPK activation in EHE cells unexpectedly increased the TAZ-CAMTA1 expression and activity. Notably, elevated TAZ-CAMTA1 expression was also associated with reduced EHE cell growth, suggesting that the induction of TAZ-CAMTA1 may be one mechanism by which AMPK suppresses EHE growth. Additionally, we found that AMPK inhibits mTOR activity and that direct mTOR inhibition also suppresses EHE cell growth. Conclusions: Together, these findings demonstrate that AMPK activation impairs EHE viability through dual mechanisms: by promoting TAZ-CAMTA1 expression and by inhibiting mTOR signaling. This work highlights AMPK as a potential therapeutic target in EHE and supports the growing body of evidence favoring mTOR inhibitors as promising treatments for this rare cancer.

Ki67 staining and single-cell RNA sequencing

Immunohistochemistry of EHE allograft tumors

Gross and microscopic pathology of lymphoma samples

To date, studies of the role for epidermal integrin a3(31 in cutaneous wound re-epithelialization have produced conflicting results: wound studies in skin from global a3-null neonatal mice have implicated the integrin in promoting timely wound re-epithelialization, whereas studies in adult mice with constitutive, epidermalspecific a3(31 deletion have not. The objective of this study was to utilize a model of inducible a3(31 deletion in the epidermis to clarify the role of a3(31 in the healing of adult wounds. We utilized the recently developed transgenic K14Cre-ERT::a3flx/flx mice (ie, inducible a3 epidermal knockout), permitting us to delete floxed Itga3 alleles (a3flx/flx) from epidermis just prior to wounding with topical treatment of 4-hydroxytamoxifen. This allows for the elucidation of a3(31-dependent wound healing in adult skin, free from compensatory mechanisms that may occur after embryonic deletion of epidermal a3(31 in the widely used constitutive a3(31-knockout mouse. We found that re-epithelializing wound gaps are larger in inducible a3 epidermal knockout mice than in control mice, indicating delayed healing, and that epidermal integrin a3(31 promotes healing of wounds, at least in part by enhancing keratinocyte proliferation. This work provides essential rationale for future studies to investigate integrin a3(31 as a therapeutic target to facilitate wound healing.

Supplemental Table 3. Single cell RNA sequencing of GEMM EHE tumors Differentially expressed genes between normal endothelial cells and Cdkn2a WT and KO tumors by single cell RNA sequencing. Functional annotations with the gene ontology sets are displayed for the Cdkn2a KO vs WT comparison, graphically displayed in Figure 3D. Transcripts used for the EHE tumor gene set are similarly displayed.

Supplemental Table 4. Bulk RNA sequencing of EHE cell lines Differential gene expression between EHE cell lines and endothelial cells via DESEQ2. Overlap in transcripts that are overexpressed by EHE cells (LogFC2>2 and FDR<0.05) are listed. These overlaps were used to generate the Venn diagram in Figure 4G.

Supplemental Table 1. Mouse Short Tandem Repeat (STR) loci for all three cell lines Short tandem repeat analysis of each EHE cell line and using both mouse and human loci. Human defining loci were undetected in all three cell lines.

Targeted therapies are effective cancer treatments when accompanied by accurate diagnostic tests that can help identify patients that will respond to those therapies. The YAP/TAZ-TEAD axis is activated and plays a causal role in several cancer types, and TEAD inhibitors are currently in early-phase clinical trials in cancer patients. However, a lack of a reliable way to identify tumors with YAP/TAZ-TEAD activation for most cancer types makes it difficult to determine which tumors will be susceptible to TEAD inhibitors. Here, we used a combination of RNA-seq and bioinformatic analysis of metastatic melanoma cells to develop a YAP/TAZ gene signature. We found that the genes in this signature are TEAD-dependent in several melanoma cell lines, and that their expression strongly correlates with YAP/TAZ activation in human melanomas. Using DepMap dependency data, we found that this YAP/TAZ signature was predictive of melanoma cell dependence upon YAP/TAZ or TEADs. Importantly, this was not limited to melanoma because this signature was also predictive when tested on a panel of over 1000 cancer cell lines representing numerous distinct cancer types. Our results suggest that YAP/TAZ gene signatures like ours may be effective tools to predict tumor cell dependence upon YAP/TAZ-TEAD, and thus potentially provide a means to identify patients likely to benefit from TEAD inhibitors.

Transient early endosome (EE)-mitochondria interactions can mediate mitochondrial iron translocation, but the associated mechanisms are still elusive. We showed that Divalent Metal Transporter 1 (DMT1) sustains mitochondrial iron translocation via EE-mitochondria interactions in triple-negative MDA-MB-231, but not in luminal A T47D breast cancer cells. DMT1 silencing increases labile iron pool (LIP) levels and activates PINK1/Parkin-dependent mitophagy in MDA-MB-231 cells. Mitochondrial bioenergetics and the iron-associated protein profile were altered by DMT1 silencing and rescued by DMT1 re-expression. Transcriptomic profiles upon DMT1 silencing are strikingly different between 2D and 3D culture conditions, suggesting that the environment context is crucial for the DMT1 knockout phenotype observed in MDA-MB-231 cells. Lastly, in vivo lung metastasis assay revealed that DMT1 silencing promoted the outgrowth of lung metastatic nodules in both human and murine models of triple-negative breast cancer cells. These findings reveal a DMT1‐dependent pathway connecting EE-mitochondria interactions to mitochondrial iron translocation and metastatic fitness of breast cancer cells.

Abstract Targeted therapies in cancer are limited by drug tolerance. In melanoma, tolerance to MAPK pathway inhibitors is associated with loss of SOX10 and enhanced YAP1/TAZ-TEAD activity. We show that loss of SOX10 is sufficient to up-regulate a TEAD transcriptional program with a strong dependence on TAZ. We developed a unique gene signature based on the transcriptomic changes following TAZ or YAP1 depletion. Expression of active TAZ was sufficient to mediate tolerance to BRAF inhibitors and MEK inhibitors. In WM983B cells, depletion of TAZ significantly reduced cell growth, whereas YAP1 knockdown resulted in little to no effect. These studies demonstrate that TAZ-TEAD activity plays an important role in melanoma drug tolerance and the development of acquired resistance. Citation Format: Connor A. Ott, Timothy Purwin, Manoela Tiago, Kristine Lou, Pan-Yu Chen, Somaneth Chowdhury, Glenn Mersky, Nir Hacohen, John Lamar, Claudia Capparelli, Gideon Bollag, Andrew Aplin. Targeting TAZ-TEAD in minimal residual disease enhances the duration of targeted therapy in melanoma models [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 7201.

Breast cancer cell analysis has traditionally focused on cell and intracellular organelle morphology. Recent research has demonstrated that organelle topology-based cancer cell classification is considerably more accurate when using handcrafted feature extraction and machine learning-based classifiers on fluorescent confocal microscopy images. However, feature extraction and classification through this methodology requires manual segmentation and computational organelle rendering. Herein, we employ convolutional neural networks (CNN) and Gradient-weighted Class Activation Mapping (GradCAM) for fast end-to-end classification and visual interpretation of confocal fluorescent microscopy images based on spatial organelle features. First, raw 3D images are filtered and preprocessed into 2D image patches for the CNN. To replicate feature analysis of the surface-surface contact area, marginal intermediate fusion CNN is implemented to classify each patch. GradCAM is then used post hoc to generate a representative heatmap of important areas used for classification. All relevant heatmap patches are then reconstructed based on the extraction of their respective patches to obtain an overall heatmap of the entire microscopy image. Furthermore, finer-grained heatmaps were obtained through the use of patch overlap and weighting during initial patch preprocessing. On a dataset consisting of 6 different breast cancer cell lines, this methodology resulted in a classification accuracy of 95.7% while also providing visualization of areas indicative of certain cancer cell lines. These findings demonstrate the efficacy of using deep learning and GradCAM for fast and interpretable organelle-based cancer cell classification.

The development of fluorescent proteins emitting in the near infrared (NIR) range (i.e., 650 nm-950 nm) has improved our capabilities for lifetime multiplexing and fluorescence imaging in vivo. Wavelengths in the NIR window experience reduced scattering and increased penetration depth through living tissue. Additionally, autofluorescence of cells and tissues is less prevalent in the NIR range, further improving signal to noise ratio. We performed fluorescence lifetime imaging (FLI) on breast cancer (AU565) and ovarian cancer (SKOV3) cell lines expressing the NIR fluorescent proteins (FPs), miRFP680 and emiRFP670. Confocal microscopy with time-correlated single-photon counting (TCSPC) reveals unique fluorescence decays for these NIR FPs, allowing for lifetime-based multiplexing on a single channel. Despite similar emission spectra, we were able to unmix fluorescence signals from a co-culture of SKOV3 expressing emiRFP670 and AU565 expressing miRFP680 based upon their unique fluorescence decays. We then generated 3D liquid overlay tumor spheroids using SKOV3 expressing emiRFP670 or miRFP680 for lifetime imaging via mesoscopic fluorescence molecular tomography (MFMT). 2D lifetime values and images acquired from MFMT corroborated our findings. Future investigation includes 3D light sheet mesoscopic imaging of tumor spheroids, as well as imaging of in vivo tumor xenografts expressing NIR-FPs. The long wavelengths and unique fluorescence lifetimes of emiRFP670 and miRFP680 make them ideal for multiplexed imaging, as well as for defining tumor volumes in vivo, while also leveraging the benefits of NIR imaging.

To date, studies of the role for epidermal integrin α3β1 in cutaneous wound re-epithelialization have produced conflicting results: wound studies in skin from global α3-null neonatal mice have implicated the integrin in promoting timely wound re-epithelialization, whereas studies in adult mice with constitutive, epidermal-specific α3β1 deletion have not. The objective of this study was to utilize a model of inducible α3β1 deletion in the epidermis to clarify the role of α3β1 in the healing of adult wounds. We utilized the recently developed transgenic K14Cre-ERT::α3flx/flx mice (ie, inducible α3 epidermal knockout), permitting us to delete floxed Itga3 alleles (α3flx/flx) from epidermis just prior to wounding with topical treatment of 4-hydroxytamoxifen. This allows for the elucidation of α3β1-dependent wound healing in adult skin, free from compensatory mechanisms that may occur after embryonic deletion of epidermal α3β1 in the widely used constitutive α3β1-knockout mouse. We found that re-epithelializing wound gaps are larger in inducible α3 epidermal knockout mice than in control mice, indicating delayed healing, and that epidermal integrin α3β1 promotes healing of wounds, at least in part by enhancing keratinocyte proliferation. This work provides essential rationale for future studies to investigate integrin α3β1 as a therapeutic target to facilitate wound healing.

Organelle Topology-based Cell Classification Pipeline can discriminate between metastatic cell lines with very high accuracy (>90%) using mitochondria or endosome immunolabeling, confocal microscopy and 3D rendering followed by several machine learning classifiers.

The development of wound therapy targeting integrins is hampered by inadequate understanding of integrin function in cutaneous wound healing and the wound microenvironment. Following cutaneous injury, keratinocytes migrate to restore the skin barrier, and macrophages aid in debris clearance. Thus, both keratinocytes and macrophages are critical to the coordination of tissue repair. Keratinocyte integrins have been shown to participate in this coordinated effort by regulating secreted factors, some of which crosstalk to distinct cells in the wound microenvironment. Epidermal integrin α3β1 is a receptor for laminin-332 in the cutaneous basement membrane. Here we show that wounds deficient in epidermal α3β1 express less epidermal-derived macrophage colony-stimulating factor 1 (CSF-1), the primary macrophage-stimulating growth factor. α3β1-deficient wounds also have fewer wound-proximal macrophages, suggesting that keratinocyte α3β1 may stimulate wound macrophages through the regulation of CSF-1. Indeed, using a set of immortalized keratinocytes, we demonstrate that keratinocyte-derived CSF-1 supports macrophage growth, and that α3β1 regulates Csf1 expression through Src-dependent stimulation of Yes-associated protein (YAP)-Transcriptional enhanced associate domain (TEAD)-mediated transcription. Consistently, α3β1-deficient wounds in vivo display a substantially reduced number of keratinocytes with YAP-positive nuclei. Overall, our current findings identify a novel role for epidermal integrin α3β1 in regulating the cutaneous wound microenvironment by mediating paracrine crosstalk from keratinocytes to wound macrophages, implicating α3β1 as a potential target of wound therapy.