Cell types and cell states identified in TNBC patients before and after chemotherapy, stratified by chemo-sensitive and chemo-resistant cases

ToC co-culture of MDA-MB-231 with ECM-myCAF, time length 72h, no treatment Scale bar = 100 μm.

Combining transcriptomic datasets from ToC and patients with TNBC reveals G0S2 as a candidate involved in ECM-myCAF–mediated chemoresistance. A, Volcano plot of differentially expressed genes in MDA-MB-231 after ToC coculture with ECM-myCAFs (compared with monoculture). Genes significantly upregulated in MDA-MB-231 upon ECM-myCAF coculture are shown in red, genes significantly downregulated are shown in blue, and genes with nonsignificant expression changes are shown in gray. B, Volcano plot of differentially expressed genes in patients with chemosensitive vs. chemoresistant TNBC at baseline. Genes significantly upregulated in chemoresistant patients are highlighted in red. Genes significantly upregulated in chemosensitive patients are highlighted in blue. Genes with nonsignificant expression changes are depicted in gray. C, Venn diagram showing the overlap between the genes upregulated in MDA-MB-231 cells after coculture with ECM-myCAFs in ToC and those upregulated in TNBC cells in chemoresistant compared with chemosensitive patients at baseline. D, UMAP from scRNA-seq data (n = 9,651 total cells; 3,866 ECM-myCAFs and 5,785 MDA-MB-231) showing G0S2 expression (left) and SAA1 expression (right). E, G0S2 expression (left) and SAA1 expression (right) in cancer cells of patients with TNBC. F, Heatmap of TF activities (analyzed using DoRothEA) in TNBC cells after mono- or coculture with ECM-myCAF in ToC. TFs that are known to regulate G0S2 transcription (according to the Signaling Pathways Project web knowledgebase) are depicted in red. C, Created in BioRender. Mechta-Grigoriou, F. (2025) https://BioRender.com/p69mf1k.

ToC mono-culture of primary ECM-myCAFs (from patient #10) time length 72h, 1μM Paclitaxel. Scale bar = 100 μm.

Human embryo implantation is a complex and finely regulated process relying on dynamiccross talk between maternal and fetal tissues. Understanding these interactions is essential for identifying the underlying causes of implantation failure or abnormal placentation. However, traditional endpoint analyses often miss transient or subtle cellular behaviors critical to this process. To overcome these limitations, we present a novel AI-integrated platform that enables continuous, in situ monitoring of early implantation events with high spatiotemporal resolution. Central to our approach is a cost-effective, rapidly prototyped implantation-on-chip (IOC) device that mimics the maternal-fetal interface using a 3D culture of human endometrial stromal cells and a 2D layer of trophoblast cells. By integrating time-lapse microscopy with advanced algorithms, our system allows automated tracking and quantification of trophoblast migration and invasion dynamics. To demonstrate the utility of this platform, we investigated the impact of nanoparticles (TiO2NPs and ZnONPs) on implantation-like behavior and observed significant alterations in trophoblast migration and invasion rates. Gene-expression analyses showed that nanoparticles modulate IL-6, IL-1 beta, and its receptor IL-1R in stromal cells, enhancing pro-inflammatory signaling and potentially affecting implantation-like processes. Altogether, these findings highlight the potential of our IOC system as a tool for studying implantation mechanisms and evaluating external influences on early pregnancy.

ToC mono-culture of primary ECM-myCAFs (sensitive, from patient #2), time length 72h, 2μM Doxorubicin. Scale bar = 100 μm.

Deep learning has proven to be one of the most effective methods in analyzing biological images to extract parameters fundamental for studying physiological functions and pathological conditions. In particular, when coupled with time-lapse microscopy (TLM), deep learning proves particularly effective in studying behaviors involving temporal dynamics. However, TLM videos are often affected by experimental noise and setup limitations, which can lead to inaccurate and poorly reproducible results. Taking advantage of the variational and generative capabilities of Variational Autoencoders (VAEs), we propose VAE-MOTION, a deep learning-based model for the analysis of cardiac contractile dynamics. By incorporating a temporal encoder into its architecture, our model allows the restoration of video quality by removing noise or increasing resolution, while simultaneously extracting accurate contraction-related signals from the latent space. The generation of synthetic videos allowed extensive training of VAE-MOTION, which subsequently validated on real videos from two different cardiac tissue models: 2D monolayers and 3D microtissues. VAE-MOTION was compared to two gold-standard methods in extracting contraction parameters relevant to drug efficacy or toxicity studies, demonstrating its potential for analyzing temporal dynamics in a given phenomenon or process.

G0S2 knockdown in TNBC cells abolishes the ECM-myCAF–mediated chemoprotective effect. A, Representative Western blots showing G0S2 (11 kDa) and β-actin (42 kDa) levels in MDA-MB-231 cells in the absence of or after 42 hours of ECM-myCAF transwell coculture. B, Quantification of A. G0S2 protein levels in TNBC cells in coculture with ECM-myCAFs are normalized to G0S2 levels in monoculture and are presented as the mean ± SEM. Several Western blots from n = 3 independent experiments. C and D, As in A and B but for MDA-MB-436 cells. E, Western blot showing G0S2 protein levels in MDA-MB-231 cells silenced (siG0S2) or not (siCTRL) for G0S2 at 48, 72, 96, and 120 hours. F, As in E but for MDA-MB-436 cells. G, Automated quantification of MDA-MB-231 siCTRL survival ± ECM-myCAFs ± doxorubicin (n = 2 independent experiments using two different patient-derived ECM-myCAFs; n = 3 videos analyzed per condition). H, Automated quantification of MDA-MB-231 siG0S2 survival ± ECM-myCAFs ± doxorubicin (n = 2 independent experiments using two different patient-derived ECM-myCAFs; n = 3 videos analyzed per condition). I and J, As in G and H but using MDA-MB-436 cells (n = 3 independent experiments using three different patient-derived ECM-myCAFs; n = 3 videos analyzed per condition). All data are represented as the mean ± SEM. B,P value from the Mann–Whitney test; for video analyses, statistical differences were assessed by the Wilcoxon matched-pair signed-rank test.

ToC co-culture of MDA-MB-436 and ECM-myCAFs, time length 72h, 1μM Paclitaxel. Scale bar = 100 μm.

Metascape analyses of cancer cells from TNBC patients and of cancer cells and ECM-myCAFs derived from the ToC system

ECM-myCAFs induce SRC kinase activity in TNBC cells. A, Kinome tree depicting upregulation and downregulation of kinase family activities in MDA-MB-231 cells upon 42 hours of ECM-myCAF transwell coculture (n = 2 different patient-derived ECM-myCAFs). B, Western blot showing pSRCTyr416, total SRC, G0S2, and β-actin protein levels in MDA-MB-231 cells after 42 hours of transwell culture without or with ECM-myCAFs (n = 4 different patient-derived ECM-myCAFs). C and D, Quantification of pSRC and G0S2 levels from B. pSRC and G0S2 protein levels in MDA-MB-231 in coculture with ECM-myCAF are normalized to their respective protein levels in monoculture. Values are represented as the mean ± SEM. E, Venn diagram showing the overlap between TFs predicted to be activated in MDA-MB-231 cells upon ECM-myCAF coculture in ToC and those known to positively regulate G0S2 transcription (as shown in Fig. 5F), together with TFs identified by Ji and colleagues (62) as regulated by v-SRC. F, Western blot showing pSRC, total SRC, G0S2, and β-actin protein levels at indicated time points in MDA-MB-231 cells silenced (siG0S2) or not (siCTRL) for G0S2. G, Quantification of F. pSRC level was normalized to siCTRL at 48 hours. H, Representative Western blot of MDA-MB-231 cells after treatment with increasing concentrations of the SRC family kinases inhibitor dasatinib. I, Quantification of G0S2 protein levels upon 48 hours of dasatinib (50 nmol/L) treatment normalized to DMSO control (n = 2 independent experiments). Values are represented as the mean ± SEM. J, Automated quantification of MDA-MB-231 survival ± ECM-myCAFs ± doxorubicin ± dasatinib (n = 3 independent experiments using three different patient-derived ECM-myCAFs; n = 3 videos analyzed per condition). C, D, and I, Statistical significance was assessed by the Mann–Whitney test; J,P values from two-sided Wilcoxon rank-sum test. E, Created in BioRender. Mechta-Grigoriou, F. (2025) https://BioRender.com/l3pzirw.

ToC co-culture of MDA-MB-231 cells and ECM-myCAFs, time length 72h, 2μM Doxorubicin. Cells were segmented using a Convolutional Neural Network (CNN). Red = cancer cells. Blue = ECM-myCAF. Scale bar = 100 μm.

Graph Deviation Networks (GDNs) are advanced models designed to capture relational data by explicitly identifying deviations from expected patterns within graph-structured systems. In respect to this, in this study we focus on the electrical activity of human brain organoids (hBOs), a biologically complex system generating time-series data with highly causal inter-regional interactions. To explore both the capabilities and limitations of GDNs in modeling neural dynamics under altered conditions, such as exposure to the neurotoxic agent valproic acid (VPA), we developed an AI-based platform to extract informative graph-based descriptors from multielectrode array (MEA) recordings. These signals were collected using a 16-electrode Axion Biosystems device, in four different sessions along 48h. To rigorously evaluate the sensitivity and robustness of the GDN architecture, we artificially introduced ghost spikes into randomly selected electrodes and correspondingly attenuated versions into adjacent electrodes, following a spatial grid-based distance criterion. We systematically assessed the predictive accuracy of the GDN under these perturbations, as well as the capacity of graph-derived descriptors to distinguish VPA-induced reductions in synaptic activity. Overall, this study advances the potential for applying GDN-based analysis in challenging experimental setups, including active electrical stimulation protocols that can temporally coincide with neuronal action potentials, especially in high-impedance brain organoid systems.Keywords—Graph Deviation Network, Multi Electrode Array, Spike template, Spike simulation, Human Brain Organoid

Primary ECM-myCAFs protect TNBC cell lines from death under doxorubicin and paclitaxel treatment in a 3D ToC model. A, Schematic overview of 3D ToC generation from primary breast cancer–derived ECM-myCAF and TNBC cell lines. B, UMAP showing high expression of FAP+ CAF (also referred to as CAF-S1) and ECM-myCAF gene signatures (10) in ECM-myCAF, validating their identity after 3D ToC culture assessed by scRNA-seq (n = 9,651 cells from two independent experiments using two different patient-derived ECM-myCAFs). C, Representative images of MDA-MB-231 ± ECM-myCAF ± doxorubicin or paclitaxel treatment at acquisition start (0 hours) and 72 hours after culture in ToC. Scale bar, 100 μm. White arrows, cancer cells; yellow arrows, ECM-myCAFs. D, Manual quantification of live MDA-MB-231 ± ECM-myCAF ± doxorubicin normalized to live cells at acquisition start (0 hours; n = 3 independent experiments using three different patient-derived ECM-myCAFs; n = 3 videos per condition). E, Manual quantification of MDA-MB-231 apoptosis ± ECM-myCAF ± doxorubicin (n = 3 independent experiments using three different patient-derived ECM-myCAFs; n = 3 videos per condition). F and G, Automated quantification of MDA-MB-231 apoptosis index (F) and survival (G) ± ECM-myCAF ± doxorubicin (n = 4 independent experiments using four different patient-derived ECM-myCAFs; n = 3 videos per condition). H and I, Manual quantification of MDA-MB-231 apoptosis index (H) and survival (I) ± ECM-myCAF ± paclitaxel (n = 3 independent experiments using three different patient-derived ECM-myCAFs; n = 3 videos per condition). J and K, Automated quantification of MDA-MB-436 apoptosis index (J) and survival (K) ± ECM-myCAF ± doxorubicin (n = 2 independent experiments using two different patient-derived ECM-myCAFs; n = 3 videos per condition). L and M, Manual quantification of MDA-MB-436 apoptotic index (L) and survival (M) ± ECM-myCAF ± paclitaxel (n = 2 independent experiments using two different patient-derived ECM-myCAFs; n = 3 videos per condition). N, Representative images displaying typical ECM-myCAF and CAP fibroblast morphology. O and P, Automated quantification of MDA-MB-231 apoptosis index (O) and survival (P) ± CAP ± doxorubicin (n = 3 videos per condition). All data are represented as the mean ± SEM. Statistical differences were assessed by the Wilcoxon matched-pair signed-rank test. A, Created in BioRender. Mechta-Grigoriou, F. (2025) https://BioRender.com/fcjaht2.

ToC co-culture of MDA-MB-436 and ECM-myCAF, time length 72h, no treatment. Scale bar = 100 μm.

ToC mono-culture of MDA-MB-231, time length 72h, no treatment. Cancer cells are stained in red, cells turning green indicate apoptotic cells. Scale bar = 100 μm.

Cell-type distribution before and after chemotherapy in chemosensitive and chemoresistant patients with TNBC. A, Flowchart illustrating the patient cohort, sample acquisition, data processing, and analysis. BC, breast cancer. B, Heatmap showing the cell-type proportions in pre- and post-chemotherapy tumor samples (N = 114 samples) from the TNBC SCANDARE Curie cohort, as inferred by deconvolution of bulk RNA-seq data using a scRNA-seq–based atlas derived from Croizer and colleagues (28). Left, N = 62 samples from 52 chemosensitive patients. Right, N = 52 samples from 36 chemoresistant patients. The response to treatment was assessed at surgery by the RCB index. Patients with an RCB index of 0 or 1 were categorized as chemosensitive, whereas those with an RCB index of II to III were defined as chemoresistant. Hierarchical clustering was applied to both samples and cell types using Euclidean distance and the ward.D2 method. Values were centered and scaled by cell types. C, Distribution of normal epithelial, cancer, fibroblastic, endothelial, and immune cells among all cell types according to chemotherapy response (N = 114 samples). D, Percentages of normal epithelial, cancer, fibroblastic, endothelial, and immune cells among total cells according to chemotherapy response (N = 114). E, Distribution of normal fibroblasts, FAP+ SMA+ CAF, and FAP− SMA+ CAF among total cells in chemosensitive and chemoresistant patients before and after treatment (N = 114). F, Distribution of FAP+ CAF clusters among all FAP+ CAF according to chemotherapy response (N = 114). G, Percentages of the different FAP+ CAF clusters among FAP+ CAF in chemosensitive and chemoresistant patients before and after chemotherapy (N = 114). Each column represents an individual patient. H–J, Percentages of cancer cells (H), ECM-myCAF (I), and detoxification iCAF (J) among total cells in chemosensitive and chemoresistant patients before and after chemotherapy (N = 114). C,P value from the Pearson χ2 test; E and H–J,P values from two-sided Wilcoxon rank-sum test. A, Created in BioRender. Mechta-Grigoriou, F. (2025) https://BioRender.com/oxvjiu9.

ToC mono-culture of MDA-MB-436, time length 72h, no treatment. Scale bar = 100 μm.

“Kiss of life” from ECM-myCAFs protects TNBC cells from doxorubicin. A, Representative image of MDA-MB-231 (red) and ECM-myCAF (blue) automatically detected. Scale bar, 100 μm. B, Measurements assessed by the algorithm: number of contacts (ECM-myCAF at minimal distance <18 μm from the centroid of the TNBC cell, Ci); minimal distance between individual TNBC cells and ECM-myCAFs; and area covered by ECM-myCAFs in a circle with a radius of 82 μm around the TNBC cell (IOU). C, Number of contacts at early time points (from treatment start to 8 hours after treatment), with respect to cancer cell fate (dead or alive; n = 92–135 measurements). D, Minimal distance at early time points, with respect to cancer cell fate (dead or alive; n = 1,109–1,420 measurements). E, IOU in the time interval at early time points, with respect to cancer cell fate (dead or alive at 72 hours after treatment; n = 795–1,341 measurements). Data are represented as individual values, including mean ± SEM. Statistical differences were assessed by the Kolmogorov–Smirnov test. B, Created in BioRender. Mechta-Grigoriou, F. (2025) https://BioRender.com/i42a245.

Identity of ECM-myCAFs before and after ToC culture and their chemoprotective effect on MDA-MB-436 cells