Cell differentiation based on cell dynamic image machine learning

Abstract

The invention relates to the field of biological medicine. Specifically, the invention relates to a cell differentiation method based on cell dynamic image machine learning. More specifically, the present invention relates to a method and apparatus for assisting in obtaining differentiated target cells (e.g., myocardial cells) from starting cells, such as pluripotent stem cells (e.g., induced pluripotent stem cells), using machine learning of cell dynamic images.

Description

Cell differentiation based on machine learning of cell dynamic images Technical Field

[0001] The present invention relates to the field of biomedicine. Specifically, it relates to a cell differentiation method based on machine learning of dynamic cell images. More specifically, it relates to a method and apparatus for utilizing machine learning of dynamic cell images to assist in obtaining differentiated target cells (e.g., cardiomyocytes) from starting cells, such as pluripotent stem cells (e.g., induced pluripotent stem cells).

[0002] Background of the Invention

[0003] Derivation of differentiated, functional cells from induced pluripotent stem cells (iPSCs) theoretically provides an unlimited source of cells for regenerative medicine, in vitro modeling of biological development and disease, and drug screening and evaluation. However, one of the major challenges of iPSC differentiation is the variability between cell lines and batches, which can lead to cells tending towards incorrect differentiation trajectories. This variability in iPSC differentiation necessitates repeated experiments, making the acquisition of functional cells time-consuming and laborious. Repeated evaluation of differentiation outcomes often relies on low-throughput or destructive methods (such as immunofluorescence), hindering quality control during the differentiation process and downstream applications. All of this severely hinders the progress of scientific research and the manufacture of cell-based products. Variation between cell lines is primarily driven by genetic and epigenetic variations in iPSCs, which can disrupt the pluripotency network and alter signaling responses in developmental pathways, resulting in varying differentiation capacities among cell lines. Other unavoidable non-genetic variations in conventional cell culture, such as cell passage number and variations in cell handling between laboratories or individuals, also contribute to differentiation variability. Furthermore, because iPSC differentiation is a stepwise process involving multiple induction stages, small perturbations or inconsistencies in early stages can accumulate and amplify, exacerbating differentiation fragility. Therefore, non-invasive monitoring and intervention of the entire differentiation process are necessary for sustained and efficient iPSC differentiation.

[0004] Currently, variability in iPSC differentiation can be partially controlled by individual experience. Based on observations of cell images, researchers can adjust experimental protocols and predict differentiation outcomes. However, this experience with cell images varies widely, making it difficult to quantify, replicate, and transfer knowledge. Furthermore, rapid or subtle changes in cell images can be difficult for researchers to detect.

[0005] Today, state-of-the-art microscopy techniques enable long-term, time-lapse, and high-throughput image acquisition of living cells. The rapidly developing field of machine learning (ML) is increasingly being applied to cell image analysis, which provides the possibility of identifying specific cellular components or cell lineages during differentiation in cell culture. During iPSC differentiation, cell fate transitions involve rapid changes in cell morphology and arrangement. Therefore, we hypothesized that microscopic images of unlabeled cells contain sufficient differentiation state information that can be captured by ML. This information can be used to intervene in the differentiation process, correct cell trajectories in a timely manner, and eliminate contamination from incorrectly differentiated cells. In this study, based on bright-field images of living cells, we developed a strategy using different ML models to non-invasively identify cell lineages, regulate the differentiation process in real time, optimize differentiation protocols, and improve the robustness of iPSC differentiation into functional cells.

[0006] BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1. The human stem cell-to-cardiomyocyte differentiation process used in this experiment. The differentiation process is divided into four stages: hiPSC differentiation, stage 1 differentiation into mesoderm, stage 2 differentiation into cardiac progenitors, and stage 3 differentiation into cardiomyocytes. This process is primarily accomplished using WNT signaling activators (CHIR) and inhibitors (IWR1). The colored arrows indicate that the duration and concentration of the small molecule CHIR application during stage 1 significantly impact differentiation efficiency.

[0008] Figure 2. Characterization of hiPSC-CMs. (a) Immunofluorescence characterization of iPSC-CMs on day 12 after differentiation; red represents cTNT, green represents MEF2C, and blue represents the nuclear dye Hoechst. (b) Immunofluorescence characterization of iPSC-CMs on day 12; green represents α-actin, blue represents the nuclear dye Hoechst. The magnified image in the white box shows distinct sarcomere structure. Scale bars are 100 μm. (c) Representative recordings of spontaneous action potentials from individual iPSC-CMs on day 15. Resting potential (Vm), frequency (f), action potential amplitude, and action potential duration at 50% amplitude (APD50) and 90% amplitude (APD90) are summarized. Data are mean ± SD. n = 4. (d) qPCR characterization of cardiomyogenic gene expression in purified iPSC-CMs compared to iPSCs.

[0009] Figure 3. Instability between cell lines and batches of hiPSC or hESC differentiation into cardiomyocytes. (a) Different cell lines have different optimal differentiation conditions, with varying optimal CHIR concentrations and ranges. Heatmap colors represent the percentage of cTnT-positive cells in different hiPSC and hESC lines treated with different CHIR concentrations on day 12 (CHIR treatment was uniform for 24 hours). (b) iPS18 cells exhibit instability across different differentiation batches under identical conditions (CHIR 6 μM for 24 hours). Green represents cTnT immunofluorescence staining. Scale bar, 1 mm.

[0010] Figure 4. Time-lapse image stream of the entire cardiomyocyte differentiation process. Live-cell brightfield image stream of hiPSC-to-cardiomyocyte differentiation and corresponding cTNT immunofluorescence staining, captured by CD7 and stitched into a full-well image (24-well plate). Scale bar: 4 mm.

[0011] Figure 5. Unsupervised clustering results of brightfield images from the entire hiPSC-CM differentiation process. (a) PCA analysis of local features of wells with low efficiency (normalized differentiation efficiency index <50%, n = 14,768) and high efficiency (normalized differentiation efficiency index ≥50%, n = 5,200). Each point in the PCA plot represents the brightfield features of a well at a specific time point during differentiation, with the different colors of the points representing different stages of differentiation. This experiment used 96-well plates, and the differentiation process lasted for 14 days, with the same field of view captured at 70-min intervals. (b) PCA analysis of brightfield images at different stages, with the color of the points representing the normalized differentiation efficiency index (%). (c)(d) LDA results of brightfield image features of wells with low dose, optimal dose, and high dose.

[0012] Figure 6. Example of a typical brightfield image at the hiPSC-CM stage. Brightfield images show a certain degree of differentiation between successful and failed differentiation. Scale bar: 0.25 mm.

[0013] Figure 7. Schematic diagram of the framework for predicting cTNT fluorescence images from brightfield images of the third stage (hiPSC-CM stage). Based on the trained model, the input brightfield image is first cropped into tiles (there is overlap between tiles, but it is not shown here for better visualization). The input tiles are first classified by GoogLeNet as "1" (positive, typically areas with a high number of hiPSC-CMs) or "0" (negative, areas with few or no hiPSC-CMs). They are then converted into fluorescence tiles using CycleGAN-1 and CycleGAN-0, respectively. These predicted tiles are then pieced back together to obtain the final predicted cTNT fluorescence image.

[0014] Figure 8. Network framework of the image classification module (GoogLeNe) and the bright field image to fluorescence image conversion module (CycleGAN). The binary classification of the image is completed by GoogLeNet, and then the images marked as "1" or "0" are converted into fluorescence images by CycleGAN-1 or CycleGAN-0 respectively; the bottom of the figure outlines the detailed architecture of CycleGAN-1, and CycleGAN-0 is not shown again because it shares the same structure with CycleGAN-1; the target generator GX→Y is trained together with a reverse generator GY→X and two discriminators DX and DY. Among them, the original CycleGAN is modified and a new "similarity loss" is added to the training target, which is expressed as

[0015] Figure 9. Accurate prediction results for brightfield to cTNT fluorescence images in the hiPSC-CM stage. (a) Typical results of CycleGAN-1 prediction for "1" patches in the brightfield image test set. Each row, from left to right, represents a single field of view: a live cell brightfield patch containing cTNT-positive hiPSC-CMs, a ground-truth cTNT immunofluorescence result, and a CycleGAN-1-predicted cTNT immunofluorescence result. The scale bar is 250μm. (b) Typical results of CycleGAN-1 prediction for "0" patches in the brightfield image test set. Each row, from left to right, represents a single field of view: a live cell brightfield patch containing almost no cTNT-positive hiPSC-CMs, a ground-truth cTNT immunofluorescence result, and a CycleGAN-0-predicted cTNT immunofluorescence result. The scale bar is 250μm. (c) Results of CycleGAN-1 and CycleGAN-0 for converting complete brightfield images to fluorescence images. Each row, from left to right, represents a single field of view, representing: a brightfield image of live hiPSC-CMs at stage 3 differentiation, the actual cTNT immunofluorescence result, and the cTNT immunofluorescence result after stitching together the predicted image tiles. Scale bar: 1 mm. (d) Comparison of the actual and predicted differentiation rates for all 36 complete brightfield images in the prediction set, measured by the Differentiation Index. (e) In the results of Figure (d), the correlation coefficient between the actual and predicted differentiation rates of the cTNT immunofluorescence images is r = 0.91 (****p < 0.0001, n = 36).

[0016] Figure 10. Schematic diagram of the framework for predicting cTnT fluorescence images from brightfield images of stage 3 (hiPSC-CMs). The pix2pix model is trained using paired brightfield and fluorescence images. The trained model can predict fluorescence labels for new brightfield images. To evaluate model performance, the model's predictions are compared with real cTnT fluorescence images.

[0017] Figure 11. Accurate prediction results for brightfield to cTNT fluorescence images at the hiPSC-CM stage. (a) Typical results of pix2pix prediction for brightfield images containing CMs. Each row, from left to right, shows: a live cell brightfield image containing cTNT-positive hiPSC-CMs, the actual cTNT immunofluorescence result, and the predicted cTNT immunofluorescence result. The scale bar is 250μm. (b) Typical results of pix2pix prediction for an image containing few CMs. Each row, from left to right, shows: a live cell brightfield image containing few cTNT-positive hiPSC-CMs, the actual cTNT immunofluorescence result, and the predicted cTNT immunofluorescence result. The scale bar is 250μm. (c) Results of pix2pix conversion of complete brightfield images to fluorescence results. Each row, from left to right, shows: a live cell brightfield image of hiPSC-CMs at the third differentiation stage, the actual cTNT immunofluorescence result, and the predicted cTNT immunofluorescence result. Scale bar is 1 mm. (d) Comparison of the true and predicted differentiation rates for all 36 complete brightfield images in the test set, measured by the Differentiation Index. (e) In the results of Figure (d), the correlation coefficient between the true and predicted differentiation rates of cTNT immunofluorescence images was r = 0.93 (****p < 0.0001, n = 36).

[0018] Figure 12. Accurate prediction of cTNT fluorescence images from brightfield images of hiPSC-CMs at the third stage of differentiation for novel cell lines. (a) pix2pix results for converting brightfield images of a complete batch of novel cell lines into fluorescence images. Each row, from left to right, represents: a brightfield image of live hiPSC-CMs at the third stage of differentiation, the actual cTNT immunofluorescence result, and the predicted cTNT immunofluorescence result. Scale bar: 1 mm. (b) Comparison of the actual and predicted differentiation efficiency indices for the novel cell line test set. Pearson correlation coefficient r = 0.81 (****p < 0.0001, n = 62).

[0019] Figure 13. Example of a typical brightfield image of hiPSC-CPC cells during differentiation. Brightfield images of successfully and unsuccessfully differentiated hiPSC-CPC cells show a certain degree of differentiation during the second differentiation stage. Scale bar: 0.25 mm.

[0020] Figure 14. A population of hiPSC-CPC cells with distinctive textures ultimately differentiated successfully. A continuous stream of brightfield images from a single field of view from differentiation day 5 to the final differentiation result. On day 6, hiPSC-CPC cells with distinctive textures in brightfield finally differentiated into cTNT-positive hiPSC-CMs, while non-CPC cells without textures in brightfield on day 6 ultimately failed to differentiate successfully. Scale bar: 0.5 mm.

[0021] Figure 15. Flowchart for weakly supervised learning-assisted prediction of differentiation efficiency at the hiPSC-CPC stage. This framework requires a trained ResNeSt-101 model to predict the presence of CPC regions capable of differentiating into CMs. Grad-CAM is used to generate a localization map for a given brightfield image when performing classification using the trained ResNeSt-101. The predicted CPC regions capable of differentiating into CMs are then obtained by binarizing the localization map. Finally, the weakly supervised learning framework is evaluated using a day 6 mask image (Grad-CAM localization map) corresponding to the input brightfield image and cTNT fluorescence images of the hiPSC-CM stage.

[0022] Figure 16. Schematic diagram of the training and testing process of the weakly supervised learning framework. During the training phase, this experiment trained the ResNeSt-101 network to classify brightfield patches. The brightfield images and corresponding mask images in the training set were cut into small blocks to obtain the dataset used to train ResNeSt-101. These mask patches contain black areas (unable to be differentiated into CM), light gray areas (uncertain whether they can be successfully differentiated into CM), and dark gray areas (successfully differentiated into CM). Based on the proportion of dark gray areas in the mask patches, we labeled the corresponding brightfield patches as "1" (positive) or "0" (negative), and discarded patches with uncertain labels. During the testing phase, in order to predict the CPC areas that can be differentiated into CM in the test set images, we first used the above-trained classification network to predict the category of the brightfield patches. For brightfield patches predicted as positive, this paper applies Gradient Weighted Class Activation Mapping (Grad-CAM) to determine the region of interest when the network predicts the brightfield patch as positive. For patches predicted as negative, the brightfield patch predictions are simply set to zero. Finally, the tile-level CPC localization map and the corresponding binary map are reassembled to obtain the complete prediction for subsequent evaluation.

[0023] Figure 17. The training process of the weakly supervised learning framework performs normally. (a) Training loss curve and validation loss curve of ResNeSt-101; (b) Classification AUC and ACC curves of ResNeSt-101.

[0024] Figure 18. Weakly supervised learning accurately predicts brightfield images of hiPSC-CPC stages. (a) Typical prediction results for a tile labeled "1" from the test set using the weakly supervised learning framework. Each row, from left to right, represents: a live cell brightfield tile from the hiPSC-CPC stage on day 6, a manually annotated mask tile, a localization tile generated using Grad-CAM, a binary tile generated from the localization tile, and cTNT immunofluorescence results on day 12. (b) Typical prediction results for a tile labeled "0" from the test set using the weakly supervised learning framework. Each row, from left to right, represents: a live cell brightfield tile from the hiPSC-CPC stage on day 6, a manually annotated mask tile, a localization tile generated using Grad-CAM, a binary tile generated from the localization tile, and cTNT immunofluorescence results on day 12. Scale bar: 250 μm.

[0025] Figure 19. Weakly supervised learning achieves excellent quantitative prediction results for brightfield images of hiPSC-CPC stages. (a) Typical prediction results of a complete hiPSC-CPC image using the weakly supervised learning framework. Each row, from left to right, represents: a live cell brightfield image of a day 6 hiPSC-CPC stage, a manually annotated mask image, a Grad-CAM localization map, a binary map of the Grad-CAM localization map, and cTNT immunofluorescence results. Scale bar is 1 mm. (b) Detailed evaluation metrics are shown in the table. The weakly supervised learning framework demonstrates superior performance. Evaluation metrics include accuracy, F1 coefficient, precision, recall, specificity, and intersection over union. (c) A visual comparison of the true differentiation index (DI) and the differentiation efficiency predicted from hiPSC-CPC images using the weakly supervised learning framework, n = 17. (d) In Figure (c), the correlation coefficient between the true differentiation index and the differentiation efficiency predicted from hiPSC-CPC images using the weakly supervised learning framework is r = 0.88 (****p < 0.0001, n = 17). e) Typical prediction results of the weakly supervised learning framework for complete hiPSC-CPC images on novel cell lines. Each row, from left to right, represents: a live cell brightfield image of a day 6 hiPSC-CPC stage, a manually annotated mask image, a Grad-CAM localization map, a binary image of the Grad-CAM localization map, and cTNT immunofluorescence results. Scale bar = 1 mm. (b) Comparison of predicted and true differentiation efficiency on novel cell lines. n = 103 wells.

[0026] Figure 20. Experimental design for DACT-1 photoactivation and (a) Flowchart for AI-CPC differentiated to day 6 using the photoactivatable small molecule DACT-1 combined with FACS purification. (b) CPC and CM can be photoactivatedly labeled and demonstrated under a microscope using laser selective area scanning. We manually selected the area to be photoactivated using brightfield images and scanned the cells within the area using a 405nm laser. The blue area in the figure represents the selected area, and the colored horizontal line represents the 405nm laser scanning trajectory. Cells within the DACT-1-labeled area can be detected in the 561nm channel. Images from left to right show: brightfield, brightfield encircled area, 561nm channel, overlay of brightfield and 561nm channel selected areas, and overlay of brightfield encircled area and 561nm channel selected area, demonstrating the accuracy of photoactivated fluorescent labeling. Scale bar: 100μm.

[0027] Figure 21. Effects of laser-assisted imaging for the purification of AI-CPCs and AI-CMs. (a) Purification of AI-CPCs on day 6 of differentiation. Immunofluorescence images of unpurified cells (CTL), differentiated cells derived from non-AI-CPCs labeled with DACT-1, and differentiated cells derived from AI-CPCs not labeled with DACT-1. Green represents cTNT and blue represents Hoechst. All cells were from the same batch and used the same differentiation conditions. After photoactivation and FACS, they were cultured in RPMI+B27 medium for another 3 days. Scale bar, 100 μm. (b) Quantification of the ratio of cTNT-positive cells in (a), n = 5. (c) Purification of AI-CPCs on day 6 of differentiation. Immunofluorescence images of unpurified cells, differentiated cells from non-AI-CPCs without DACT-1 labeling, and differentiated cells from AI-CPCs labeled with DACT-1. cTNTs are shown in green and Hoechst is in blue. All cells were from the same batch and used the same differentiation conditions. After photoactivation and FACS, they were cultured in RPMI+B27 medium for another 3 days. Scale bars represent 100 μm. (d) Quantification of the cTNT-positive cell ratio in panel (c), n = 5. (e) Purified CMs from day 12 of differentiation. Immunofluorescence images of unpurified cells, non-CMs labeled with DACT-1, and CMs not labeled with DACT-1. cTNTs are shown in green and Hoechst is in blue. All cells were from the same batch and used the same differentiation conditions. After photoactivation and FACS, they were cultured in RPMI+B27 medium for another 3 days. Scale bars represent 100 μm. (f) Quantification of the cTNT-positive cell ratio in panel (e), n = 5. * indicates p < 0.05, **** indicates p < 0.0001. Statistical analysis of the above figures was performed using one-way ANOVA followed by Dunnett's multiple comparison test.

[0028] Figure 22. Immunofluorescence analysis of AI-CPCs reveals that they possess essential characteristics of cardiomyocyte progenitor cells. (a) Brightfield image of AI-CPCs on day 6 of differentiation. Regions with textured features show widespread expression of CPC-specific genes, including GATA4, MEF2C, NKX2.5, and ISL1, while non-AI-CPC regions show slightly weaker fluorescence signals. Under highly efficient differentiation conditions, a small number of AI-CPCs may stain weakly with cTNT-positive signals. Scale bar: 20 μm. (b) Quantification of panel (a), n = 5.

[0029] Figure 23. The expression profile of AI-CPC exhibits CPC characteristics. (a) PCA analysis results of bulk RNA-seq. The horizontal axis is the first principal component (70.6%), and the vertical axis is the second principal component (19.1%). Each point represents an RNA-Seq sample, n = 3. (b) Whole-genome gene expression heatmap of hiPSC, AI-CPC, non-AI-CPC, and CM. Gene expression levels were normalized between samples using Log2 (FPKM+1). A total of 17,561 genes were analyzed by hierarchical clustering. (c) Partial gene expression heatmap of hiPSC, AI-CPC, hiPSC-CM, and non-CPC, including five independent gene types, from top to bottom: pluripotency genes, endothelial cell-specific genes, fibroblast or epicardium-specific genes, CPC, and CM-related genes; gene expression levels were normalized between samples using Log2 (FPKM+1). (d) GO analysis using the top 500 differentially expressed genes (DEGs) compared between AI-CPC and hiPSCs revealed enrichment in the top 20 genes, most of which were related to heart or cardiomyocyte development. (e) GO analysis using the top 500 differentially expressed genes (DEGs) compared between non-AI-CPC and hiPSCs revealed enrichment in the top 20 genes.

[0030] Figure 24. Discovery of differentiation patterns at the edge and center of stem cell clones. (a) Brightfield images and cTNT staining results from the 0h stem cell stage to the final differentiation stage in a unified field of view. To more clearly show the edge of the cell clone, the brightfield images were enhanced. Scale bar: 2 mm. (b) Merged 24-hour live cell brightfield image and day 12 hiPSC-CM image of the same region with cTNT staining. It can be seen that gaps without hiPSC coverage in the first stage are more likely to successfully differentiate into cardiomyocytes. Scale bar: 500 μm. (c) Quantification of the percentage of hiPSCs or hiPSC-free areas (24-hour images) in cTNT-positive areas (day 12 images) between the same well and randomly different wells (CTL).

[0031] Figure 25. (a) Machine learning evaluates differentiation efficiency based on starting clone status. At the iPSC stage, image features are passed to a random forest model, which then predicts differentiation efficiency, providing guidance for selecting the optimal differentiation starting point. (b) Feature importance ranking results determined using the random forest model. (c) PCA plot of 343 features. Each dot represents a well, and the color of the dot indicates its final differentiation efficiency index. (d) Relationship between the values ​​of the 8 most important features in (b) and the final differentiation efficiency. (e) Predicted and actual differentiation efficiency indices by the random forest model. Test set n = 584 wells.

[0032] Figure 26. Clone size significantly affects differentiation efficiency. (a) Brightfield images of hiPSC clones of varying sizes. Clone size was controlled by enzyme digestion time and manipulation during passaging, ensuring the same initial number of hiPSC cells per well; scale bar, 200 μm. (b) Bar graph showing the effect of different starting hiPSC clone sizes on differentiation efficiency (cTNT-positive cells), using RPMI+B27 and RPMI+S12 basal media, respectively.

[0033] Figure 27. The optimal CHIR treatment concentration and duration during the first differentiation phase show a negative correlation. Different CHIR concentrations and durations during the first differentiation phase significantly impact the proportion of cTNT-positive hiPSC-CMs at the end of differentiation. The horizontal axis represents the actual CHIR concentration, and the vertical axis represents the duration of CHIR treatment (CHIR treatment duration does not affect the timing of IWR1 addition; IWR1 is uniformly added at 72 hours). The color of the scattered dots represents the final differentiation efficiency.

[0034] Figure 28. Switching the appropriate CHIR concentration every 24 hours during the first stage can still improve differentiation efficiency. (a) Using a single CHIR concentration for 0-24 hours of differentiation, and switching the CHIR concentration between 24-48 hours, differentiation efficiency can be rescued by adjusting the CHIR concentration in the second half of differentiation. (b) Using a single CHIR concentration for 0-24 hours of differentiation, and switching the CHIR concentration between 24-32 hours, differentiation efficiency can be rescued by adjusting the CHIR concentration in the second half of differentiation. Point color represents final differentiation efficiency.

[0035] Figure 29. Schematic diagram of the working principle and brightfield feature extraction and analysis model for determining the relative CHIR concentration in the first stage of differentiation. (a) Workflow of the brightfield image classification system for the first stage of myocardial differentiation. A live cell brightfield image stream of a well taken between 0 and 12 hours is input. The classification system is required to predict whether the CHIR concentration is low, moderate, or high. Specifically, relevant image features are first extracted from the input image stream, and then a machine learning classifier is used to make a prediction based on these features. Wells predicted to have low or high CHIR concentrations can be rescued by adjusting their CHIR dose to further stabilize the differentiation system. (b) Schematic diagram of the classification system training. The training dataset contains brightfield image streams of many wells and their corresponding concentration labels. These wells are mapped into points in a high-dimensional feature space. During training, the logistic regression classifier aims to linearize the decision boundary to maximize the separation of points of different classes. (c) Schematic diagram of feature extraction from brightfield images taken between 0 and 12 hours. The image stream consists of 10 images uniformly captured between 0 and 12 hours. There are two types of features: Type I features are calculated at each timestamp; Type II features are calculated between two consecutive timestamps. Both types of features produce a list of real numbers representing the change in the feature between T1 and T10 (0-12 hours). These feature values ​​are then further processed in either an absolute or relative manner: in the "absolute" mode, the raw values ​​of the feature sequence are used; in the "relative" mode, the raw values ​​of the sequence are normalized by dividing them by the first number. Finally, the feature sequence is divided into early, middle, and late stages, and the average of the feature values ​​for each stage is calculated. Among the features designed in this paper, "local entropy," "cell brightness," and "fractal dimension" are first-type absolute features; "area," "perimeter," and "area-perimeter ratio" are first-type relative features; and "optical flow" is a second-type relative feature. Finally, each feature produces three real numbers (early, middle, and late), resulting in a 21-dimensional feature representation for each well.

[0036] Figure 30. Concentration estimation using a machine learning model. (a) LDA dimensionality reduction plot of all features. (b) Classification performance using all features. (c) PCA dimensionality reduction plot of all features. (d) PCA dimensionality reduction plot after feature filtering (selecting the four features with the highest importance weights). (e) LDA dimensionality reduction plot after feature filtering. (f) Feature importance ranking at 24 hours. (g) Classification performance after feature filtering.

[0037] Figure 31. Results of cross-batch cross-validation for CHIR concentration estimation. (a) There are four batches (denoted by CD01-1, 01-2, 01-3, and 01-4). In each round, the classification model is trained and feature-selected on three batches, and predictions are made on the remaining batches. For each concentration level used in the test batch, all wells using that concentration condition are input into the trained classifier, and their predictions are aggregated into a "deviation score" (ranging from -1 to +1). This deviation score reflects the degree to which the concentration deviates from the optimal range, providing guidance to the experimenter in determining the optimal concentration range and subsequently rescuing wells with high or low concentrations. (b) Comparison of the predicted "deviation score" and the actual "ΔCHIR concentration," as well as the Pearson correlation coefficient.

[0038] Figure 32. RNA-seq reveals differentiation toward somite mesoderm in the high-dose CHIR group. (a) PCA analysis of samples at different CHIR doses shows a relatively concentrated location of the successfully differentiated groups. During the first stage of differentiation, sequencing samples were collected after CHIR treatment for the corresponding time periods. The color of the dots represents the efficiency of hiPSC-CM differentiation under different conditions. (b) Genome-wide heatmap clustering results for samples treated with different CHIR treatment times and concentrations. Gene expression levels were normalized across samples using Log2 (FPKM+1). (c) Heatmap of selected genes shows that different CHIR doses determine different differentiation directions. Genes related to cardiac mesoderm are upregulated in the moderate-dose group, while genes related to presomitic mesoderm are significantly upregulated in the high-dose group, potentially interfering with cardiac mesoderm fate determination. Gene expression levels were normalized across samples using Log2 (FPKM+1). (d) GO analysis showed that the functions of DEG genes enriched in the high-dose group (CHIR 10 μM 48 h) were related to the development of somitogenesis and anterior / posterior patterning compared with the medium-dose group (CHIR 6 μM 48 h).

[0039] Figure 33. Knockdown of MSX1 effectively inhibits presomite mesoderm differentiation under conditions of high CHIR concentration and prolonged treatment time. (a) At the same CHIR treatment time (48 h), MSX1-knockdown hiPSCs are able to adapt to higher CHIR concentrations. (b) At the same CHIR concentration (16 μM), MSX1-knockdown hiPSCs are able to adapt to longer CHIR treatment times. Scale bar: 200 μm. (c) Differentiation efficiency of control hiPSCs and MSX1-knockdown hiPSCs (C8, C9) at different levels of WNT signaling activation. C8 and C9 represent two different shRNAs targeting the MSX1 gene, respectively.

[0040] Figure 34. Small molecule screening flow chart. (a) The goal of screening small molecules is to enable normal cardiomyocyte differentiation in the high-dose CHIR group, with differentiation efficiency predicted by bright-field imaging on day 6 as the evaluation criterion. (b) Specific drug development strategy.

[0041] Figure 35. Schematic overview of an iPSC differentiation strategy using image-based machine learning, using cardiomyocyte (CM) differentiation as an example to address efficiency variability. Top: Variation occurs at every step of the iPSC differentiation process. Bottom: Using brightfield image-based machine learning, the present strategy can be applied at different stages to reduce variation and achieve highly efficient CM induction.

[0042] Figure 36. Early assessment of CHIR concentration in early renal differentiation using machine learning. (a) Schematic diagram of iPSC-to-NPC differentiation using CHIR as an inducer. Red arrows indicate that different CHIR concentrations on days 0-4 resulted in different differentiation outcomes. NPCs were collected on day 9 for immunofluorescence staining of SIX2. (b) Representative bright-field images of cells on day 4 treated with low, optimal, and high CHIR concentrations. Scale bar, 200 μm. (c) Representative immunofluorescence images of NPCs on day 9 treated with low, optimal, and high CHIR concentrations for SIX2. Scale bar, 200 μm. (d) T-SNE of local features of bright-field images on day 4 on the training set. n = 3,398. (e) Classification performance of the logistic regression model on the test set. (f) Confusion matrix of the logistic regression model on the test set, n = 1,457.

[0043] Figure 37. Definitive endoderm identification in early liver differentiation using machine learning. (a) Schematic diagram of iPSC-to-definitive endoderm differentiation using small molecules for hepatocyte-like cell induction. DEs were collected on day 3 for SOX17 immunofluorescence staining. (b) Representative immunofluorescence results for SOX17 (green) and Hoechst (blue) on DEs on day 3. Images with different final efficiencies (proportion of SOX17+ area) were selected. Scale bar, 100 μm. (c) Representative prediction results for DE cell identification on bright field images of the test set. From left to right: bright field image of live cells on day 3; Grad-CAM heat map of endoderm cell localization; binary prediction of SOX17+ endoderm cell localization; true SOX17 fluorescence result on day 3; SOX17 fluorescence image enhanced by binarization and morphological operations. Scale bar, 1 mm. (d) Correlation between the true differentiation efficiency (derived from SOX17 fluorescence results on day 3) and the predicted differentiation efficiency (prediction based on bright field images on day 3), as well as Pearson's r value.

[0044] Figure 38. Architecture of the pix2pix model for fluorescence prediction. (a) The pix2pix model is trained for brightfield to fluorescence image conversion. The generator G learns to predict fluorescence images from brightfield images, while the discriminator D learns to distinguish between real and fake brightfield-fluorescence image pairs. (b) Detailed architecture of the generator. The generator G is a U-Net with 8 convolutional layers in both the encoder and decoder. All inner convolutional layers are followed by instance normalization and ReLU activations. The transposed convolutions in the original design are replaced with nearest neighbor upsampling followed by 5×5 convolutions. (c) Detailed architecture of the discriminator. The discriminator D is a 3-layer convolutional neural network. Each pixel in the network output has a receptive field of size 16×16, representing the real / fake classification score for the corresponding 16×16 patch.

[0045] Figure 39. Specific process for locating CPC regions using weak supervision. During the training phase, this experiment trained the ResNeSt-101 network to classify brightfield patches. The brightfield images and corresponding mask images in the training set were cut into small blocks to obtain the dataset used to train ResNeSt-101. These mask patches contain black areas (unable to be differentiated into CM), light gray areas (uncertain whether they can be successfully differentiated into CM), and dark gray areas (successfully differentiated into CM). Based on the proportion of dark gray areas in the mask patch, we labeled the corresponding brightfield patch as "1" (positive) or "0" (negative), and discarded patches with uncertain labels. During the testing phase, in order to predict the CPC regions that can be differentiated into CM in the test set images, we first used the above-trained classification network to predict the category of the brightfield patch. For brightfield patches predicted as positive, this paper applies Gradient Weighted Class Activation Mapping (Grad-CAM) to determine the region of interest when the network predicts the brightfield patch as positive. For patches predicted as negative, the brightfield patch predictions are simply set to zero. Finally, the tile-level CPC localization map and the corresponding binary map are reassembled to obtain the complete prediction for subsequent evaluation.

[0046] Detailed Description of the Invention

[0047] In one aspect, the present invention provides a neural network model for predicting the efficiency of differentiation of starting cells into target cells, which is obtained by the following steps:

[0048] A bright field image of cells at a specific stage of differentiation is provided as an input image, and a corresponding target cell image confirmed by target cell-specific staining is used as a correct image, and a neural network is used for learning to obtain the neural network model.

[0049] In some embodiments, the neural network includes (1) an image classification neural network, and (2) an image conversion neural network.

[0050] In some embodiments, the starting cell is a pluripotent stem cell, such as an embryonic stem cell (eg, an embryonic stem cell no older than 14 days) or an induced pluripotent stem cell.

[0051] In some embodiments, the target cell is a differentiated cell, for example, the cell is selected from a neuronal cell, a skeletal muscle cell, a hepatocyte, a kidney cell, a fibroblast, an osteoblast, a chondrocyte, an adipocyte, an endothelial cell, a mesenchymal cell, a smooth muscle cell, a cardiomyocyte, a neural cell, a hematopoietic cell, or a pancreatic islet cell.

[0052] In some embodiments, the (1) image classification neural network is selected from googleNet, VGG, ResNet, ResNeXt and SE-Net, preferably googleNet.

[0053] In some embodiments, the (2) image conversion neural network is selected from CycleGAN, DiscoGAN and DualGAN, preferably CycleGAN.

[0054] In some specific embodiments, the (1) image classification neural network is GoogleNet, and the (2) image conversion neural network includes two CycleGANs. In some embodiments, GoogleNet classifies patches of brightfield images into categories "0" and "1," and then inputs the patches and corresponding stained patches into CycleGAN-0 and CycleGAN-1, respectively, for learning.

[0055] In some embodiments, the neural network comprises a pix2pix model. In some embodiments, the pix2pix model comprises a generator G that learns to predict stained images from brightfield images, and a discriminator D that learns to distinguish between real and fake brightfield-fluorescence image pairs.

[0056] In some embodiments, the neural network is a random forest regression model.

[0057] In some embodiments, the following features of the bright field image are used to quantify the morphological characteristics of the cells:

[0058] (1) local entropy, cell brightness, cell contrast, and total variation;

[0059] (2) Hu invariant moments 1–7;

[0060] (3) SIFT 1-256;

[0061] (4) ORB 1~64;

[0062] (5) Area, perimeter, and area / perimeter ratio;

[0063] (6) Solidity, convexity, and roundness;

[0064] (7) Maximum center-contour distance (CCD), minimum CCD, minimum / maximum CCD ratio, mean CCD, standard deviation of CCD; and

[0065] (8) Spacing.

[0066] In some embodiments, the specific stage of differentiation is the final stage of induced differentiation.

[0067] In some embodiments, the specific stage of differentiation is an intermediate stage of induced differentiation.

[0068] In some embodiments, the specific stage of differentiation is the initial stage of induced differentiation.

[0069] In some embodiments, the cells are treated with a given condition at a specific stage of differentiation. In some embodiments, the cells are treated with a given small molecule at a specific stage of differentiation. In some embodiments, the small molecule is a small molecule that is critical for the differentiation of the cells. For cardiomyocyte differentiation, the small molecule is CHIR99021.

[0070] In some embodiments, the target cell is a cardiomyocyte.

[0071] In some embodiments, the target cell-specific staining is immunofluorescence staining.

[0072] Specific staining methods for different target cells are readily known to those skilled in the art. For example, for cardiomyocytes, immunofluorescence staining for cardiac troponin T (cTNT) can be used. For hepatocytes, immunofluorescence staining for SOX17 can be used. For kidney cells, immunofluorescence staining for SIX2 can be used. Immunofluorescence staining can be performed using commercially available kits.

[0073] In another aspect, the present invention provides a neural network model for predicting the cell regions capable of differentiating into target cells during differentiation of starting cells into target cells, which is obtained by the following steps:

[0074] A bright field image of cells at a specific stage of differentiation is provided as an input image, and a corresponding image of cells suspected of being able to differentiate into target cells is used as a correct image. A neural network is used to perform weakly supervised learning to obtain the neural network model, wherein the neural network includes (1) an image classification neural network, and (2) an image positioning neural network.

[0075] In some embodiments, the starting cell is a pluripotent stem cell, such as an embryonic stem cell or an induced pluripotent stem cell.

[0076] In some embodiments, the target cell is a differentiated cell, for example, the cell is selected from a neuronal cell, a skeletal muscle cell, a hepatocyte, a kidney cell, a fibroblast, an osteoblast, a chondrocyte, an adipocyte, an endothelial cell, a mesenchymal cell, a smooth muscle cell, a cardiomyocyte, a neural cell, a hematopoietic cell, or a pancreatic islet cell.

[0077] In some embodiments, the (1) image classification neural network is selected from Resnet-101, VGG, ResNeXt, SE-Net, preferably Resnet-101.

[0078] In some embodiments, the (2) image localization neural network is selected from Grad-CAM.

[0079] In some embodiments, the specific stage of differentiation is the final stage of induced differentiation.

[0080] In some embodiments, the specific stage of differentiation is an intermediate stage of induced differentiation.

[0081] In some embodiments, the specific stage of differentiation is the initial stage of induced differentiation.

[0082] In some embodiments, the cells are treated with a given condition at a specific stage of differentiation. In some embodiments, the cells are treated with a given small molecule at a specific stage of differentiation. In some embodiments, the small molecule is a small molecule that is critical for the differentiation of the cells. For cardiomyocyte differentiation, the small molecule is CHIR99021.

[0083] In some embodiments, the target cell is a cardiomyocyte.

[0084] In some embodiments, the target cell-specific staining is immunofluorescence staining.

[0085] In some embodiments, the specific stage of differentiation is the mesodermal cell stage.

[0086] In some embodiments, the entire bright field image is segmented into tiles, and the tiles are labeled as ground-truth labels ("0": negative, "1": positive) or uncertain labels (Uncertainlabels) according to the proportion of successfully differentiated areas in the tiles;

[0087] The ResNeSt-101 network was trained using a training dataset consisting of brightfield image patches with deterministic labels.

[0088] Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to generate localization maps to visualize the differentiable cell regions.

[0089] In another aspect, the present invention provides a method for predicting the efficiency of differentiation of a starting cell into a target cell, the method comprising:

[0090] (1) Acquire bright-field images of cells at specific stages of differentiation;

[0091] (2) analyzing the bright field image using the neural network model of the present invention for predicting the efficiency of differentiation of starting cells into target cells;

[0092] (3) Determine the differentiation efficiency.

[0093] In some embodiments, differentiation efficiency is quantified by a differentiation index (or differentiation efficiency index), wherein

[0094] For an MxN fluorescence staining image I (intensity value ∈ [0, 1]), its “differentiation efficiency index” is defined as the total fluorescence intensity of pixels whose intensity value exceeds the threshold α, that is,

[0095] Where M and N are the height and width of the fluorescence image.

[0096] In another aspect, the present invention provides a method for predicting a cell region capable of differentiating into a target cell during differentiation of a starting cell into a target cell, the method comprising:

[0097] (1) Acquire bright-field images of cells at specific stages of differentiation;

[0098] (2) analyzing the bright field image using the neural network model of the present invention for predicting the cell region capable of differentiating into target cells during the process of differentiation from starting cells to target cells;

[0099] (3) Identify the cell regions that can differentiate into target cells.

[0100] In some embodiments, the specific stage of differentiation is the final stage of induced differentiation.

[0101] In some embodiments, the specific stage of differentiation is an intermediate stage of induced differentiation.

[0102] In some embodiments, the specific stage of differentiation is the initial stage of induced differentiation.

[0103] In some embodiments, the target cell is a cardiomyocyte.

[0104] In some embodiments, the target cell-specific staining is immunofluorescence staining.

[0105] In some embodiments, the specific stage of differentiation is the mesodermal cell stage.

[0106] Based on the determined area of ​​cells capable of differentiating into target cells, the differentiation efficiency can also be predicted / determined, for example, by area ratio.

[0107] In another aspect, the present invention provides a method for isolating and / or purifying cells at a specific stage in the process of differentiation from starting cells to target cells, the method comprising:

[0108] (1) Acquire bright-field images of cells at specific stages of differentiation;

[0109] (2) analyzing the bright field image using the neural network model of the present invention for predicting the cell region capable of differentiating into target cells during the process of differentiation from starting cells to target cells;

[0110] (3) identifying regions of cells capable of differentiating into target cells;

[0111] (4) treating cells with a laser-activated probe such as DACT-1;

[0112] (5) treating cells outside the region of cells determined to be capable of differentiating into target cells with a laser, and

[0113] (6) Cells within the cell region determined to be capable of differentiating into target cells are sorted out.

[0114] In some embodiments, the sorted cells have an increased proportion of cells that differentiate into target cells.

[0115] In some embodiments, the laser-activated probe is a toxic laser-activated probe.

[0116] In some embodiments, the target cell is a cardiomyocyte and the cell at a specific stage is a cardiac progenitor cell.

[0117] In another aspect, the present invention provides a method for screening conditions that can promote differentiation of starting cells into target cells, the method comprising:

[0118] 1) Changing one or more differentiation conditions at a specific stage of differentiation;

[0119] 2) predicting / determining the differentiation efficiency under the altered differentiation conditions by the method of the present invention;

[0120] 3) Determine the conditions under which the differentiation efficiency is optimal as the conditions that promote differentiation.

[0121] In some embodiments, the differentiation condition is contact with a given small molecule compound to be tested, for example, differentiation is performed in a culture medium containing a given small molecule compound to be tested.

[0122] In some embodiments, the target cells are cardiomyocytes. In some embodiments, the specific stage of differentiation is the stage of differentiation of pluripotent stem cells into cardiac mesoderm. In some embodiments, the differentiation conditions are the addition of a small molecule compound to be tested at a given CHIR99021 concentration.

[0123] Cardiomyocyte differentiation typically involves providing iPSCs. The first stage (0-about 72 hours) involves culturing in the presence of a WNT signaling pathway activator, such as CHIR99021 (CHIR); the second stage involves culturing in the presence of a WNT signaling pathway inhibitor, such as IWR1, for approximately 48 hours; and the third stage involves adding insulin to the basal differentiation medium to allow the cells to spontaneously differentiate into beating cardiomyocytes. The entire process involves four stages: stem cells (iPSCs), cardiac mesoderm (Stage I), cardiac progenitor cells (CPCs, Stage II), and cardiomyocytes (CMs, Stage III). Typically, beating cardiomyocytes can be observed under a microscope within 7-10 days.

[0124] In another aspect, the present invention provides a method for differentiating cardiomyocytes from pluripotent stem cells, such as embryonic stem cells (e.g., embryonic stem cells no older than 14 days) or induced pluripotent stem cells, the method comprising:

[0125] 1) At the pluripotent stem cell stage (differentiation initiation stage), the method of the present invention is used to predict and / or determine the differentiation efficiency, thereby performing quality control on the starting pluripotent stem cells;

[0126] 2) At early stages of differentiation (e.g., mesoderm stage), using the methods of the present invention to predict and / or determine differentiation efficiency, thereby evaluating early differentiation conditions and maintaining or modifying the differentiation conditions accordingly;

[0127] 3) At the middle or late stages of differentiation (such as the cardiac progenitor cell CPC or cardiomyocyte CM stage), using the method of the present invention to predict and / or determine the differentiation efficiency, thereby ending the differentiation or continuing the differentiation accordingly; and / or

[0128] 4) Purifying differentiated intermediate cells capable of differentiating into cardiomyocytes based on the method of the present invention, thereby improving differentiation efficiency.

[0129] In another aspect, the present invention provides a system / device for implementing the method of the present invention. The system / device, for example, comprises at least an image acquisition module (such as a bright field image acquisition module) and a neural network module comprising the neural network model of the present invention. Example

[0130] 1. Experimental Methods

[0131] 1.1 Passaging and culture of stem cells

[0132] The hiPSCs and hESCs used in this experiment were routinely cultured in 6-well plates, passaged approximately every four days, and placed in a cell culture incubator maintained at 37°C and 5% CO2. The passage procedures are detailed below:

[0133] 1) Observe the cell density under a microscope. When the cells have proliferated to about 80% of the total area, they are ready for subculturing.

[0134] 2) Before passaging, pre-spread Matrigel in the well plate. All Matrigel operations must be performed on ice. Dilute the original Matrigel 50-fold with pre-chilled DMEM / F-12 and add enough to cover the bottom of the well plate (850 μL / well for a 6-well plate, for example). Incubate in a 37°C incubator for 30 minutes. Aspirate the liquid before use.

[0135] 3) Preheat the stem cell culture medium PGM1 or CDM (depending on the purpose of the subsequent experiment), DPBS, and Versene to 37°C in advance, and add Y27632 (5 μM) to the stem cell culture medium;

[0136] 4) After removing the cells from the incubator, aspirate the culture medium. Add 1 ml of DPBS to each well to wash the remaining culture medium, aspirate it, add 1 ml of Versene, and digest at 37°C for 3 minutes.

[0137] 5) Immediately remove the Versene and pipette the cells 3-4 times with 1 ml of PGM1 medium to release the cells from the bottom of the dish.

[0138] 6) Mix the cell suspension with the remaining culture medium and add to a new plate covered with Matrigel. The cell culture ratio should be 1:6-1:12, but this ratio may be adjusted slightly depending on the starting cell number.

[0139] 7) Replace the PGM1 medium with fresh one 12-24 hours after passage to remove the Y27632 medium. Observe the cell status and density daily.

[0140] 1.2 Cardiomyocyte differentiation

[0141] Stem cell to cardiomyocyte differentiation is routinely performed in 24-, 96-, or 384-well plates. The steps are detailed below (Figure 3.1):

[0142] 1) hiPSCs were separated into CDM medium at a ratio of 1:10 or 1:12. The separation steps were the same as the above passaging steps. Y27632 (5 μM) was added to the CDM medium. The day was recorded as -3.

[0143] 2) 12-24 hours after passage, change the medium to remove Y27632 and continue culturing in CDM medium. Observe the cell status and density daily.

[0144] 3) When hiPSCs reach 80-90% confluence, replace the culture medium with RPMI+B27minus (50 ml RPMI1640 culture medium + 1 ml B27minus. Note that the B27 additive should be stored in a -20°C refrigerator and aliquoted into small packages for use). Add 2-20 μM CHIR. This is marked as differentiation day 0, and the hiPSC stage is complete.

[0145] Note: The dosage of CHIR is flexible and unstable. Different cell lines, different batches, and different operators can lead to large differences in differentiation results.

[0146] 4) After 24-48 hours, the culture medium was replaced with RPMI+B27minus;

[0147] 5) At 72 h, the culture medium was replaced with RPMI+B27minus and 15 μM of the small molecule IWR was added. This was recorded as the third day of differentiation. At this time, the cells differentiated into the mesoderm stage, and the first stage ended.

[0148] 6) 48 hours after IWR1 addition, the culture medium was replaced with RPMI + B27 minus, and IWR1 was removed, marking differentiation day 5. 7) Cultured in RPMI + B27 minus for 24 hours, marking differentiation day 6, at which point the cells differentiated into hiPSC-CPCs. The culture medium was then replaced with RPMI + B27, marking the end of the second stage.

[0149] 8) Continue culturing in RPMI + B27 medium, changing the medium every 3 days. The cells will spontaneously differentiate into beating hiPSC-CMs within 3-6 days, marking the third stage of differentiation. Cell beating can be observed as early as day 7-8.

[0150] In addition, RPMI+S12 can also support efficient hiPSC-CM differentiation. Except for replacing B27 supplement with S12, the rest of the operation process is consistent with the above. For details, please refer to the reference S12 medium (Peietal., 2017).

[0151] 1.3 Purification of cardiomyocytes by metabolic pathways

[0152] On differentiation days 10-12, aspirate the RPMI+B27 medium used for hiPSC-CMs, wash with DPBS, and then replace with DMEM (glucose-free, glutamine-free, and phenol red-free) supplemented with 4 mM L-lactic acid as a carbon source. Refresh the medium every three days to promptly remove dead cells. After 3-6 days of continuous culture, nearly all non-cardiomyocytes will be significantly dead.

[0153] 1.4 Cardiomyocyte Digestion

[0154] The operation of the hiPSC-CMs digestion process significantly affects the state and quality of subsequent hiPSC-CMs. Using earlier beating hiPSC-CMs for better digestion results. The longer the hiPSC-CMs are cultured after successful differentiation, the more difficult it is to digest them into single cells. The detailed steps are as follows:

[0155] 1) Dilute 0.25% trypsin to 0.05% in PBS and preheat at 37°C.

[0156] 2) Take out the successfully differentiated hiPSC-CMs cells, dry the RPMI+B27 medium used, and use PBS

[0157] Wash away the residual culture medium to avoid affecting the digestion effect;

[0158] 3) After aspirating PBS, digest hiPSC-CMs with 0.05% trypsin at 37°C for 5-7 minutes, then gently shake in a 37°C water bath for 2 minutes;

[0159] 4) After gently pipetting 2-3 times to disperse the cells, filter the cells through a 40 mm pore size cell strainer and resuspend them in RPMI + B27 medium containing 10% FBS. FBS is used to neutralize the damage caused by trypsin to the cells.

[0160] 5) Centrifuge at 850 rpm for 3 minutes, remove the supernatant, count the cells, resuspend them in RPMI+B27 medium containing 10% FBS and 5 μM Y27632, and plate them into well plates;

[0161] 6) 12-24 hours after plating the cells, replace the culture medium with RPMI+B27 medium and change the medium every 3 days.

[0162] 1.5 Culture and Passaging of 293T Cells

[0163] 293T cells are used for lentiviral packaging, and their status significantly affects subsequent packaging efficiency. The detailed steps are as follows:

[0164] 1) Aspirate the culture medium from the 10 cm culture dish, add 1 ml of PBS, and gently wash the cells;

[0165] 2) Aspirate the PBS and add 1 mL of 0.25% trypsin to the culture dish. Gently shake the dish to evenly distribute the trypsin. Digest the cells in a 37°C incubator for 1 min. Observe the cells under an inverted microscope to ensure complete digestion and separation. Completely digested and separated cells should appear as round pellets.

[0166] 3) Add 1 mL of DMEM high-glucose medium to terminate the trypsin reaction;

[0167] 4) Use a pipette to draw up the culture medium and swirl it against the upper half of the culture dish at a 45° angle several times to break up the cells.

[0168] 5) Use a pipette to draw the cell-containing culture medium into a centrifuge tube and centrifuge at 1000 rpm for 3 minutes. After centrifugation, observe the bottom of the centrifuge tube to see if there are any cell deposits.

[0169] 6) Aspirate the supernatant and gently tap the bottom of the centrifuge tube to disperse the cell clumps;

[0170] 7) Add 5 mL of DMEM high-glucose medium, pipette the liquid, and pipette repeatedly to disperse the cells into single cells;

[0171] 8) Split the cells at a ratio of 1:2-1:3 and fill the dish with DMEM high-glucose medium to 10 mL / dish. Use a pipette to add to the entire dish. Gently shake the dish to evenly distribute the cells and place in an incubator for culture.

[0172] 1.6 Lentivirus Preparation and Infection

[0173] The lentiviral vector used in this experiment was modified from a lentiviral vector, with vesicular stomatitis virus G protein (VSV-G) as the envelope protein, plus the expression protein plasmid pRSVREV to help nuclear export and shell assembly, the polyprotein expression gene Gag with envelope and matrix, the polyprotein expression gene Pol of protease, reverse transcriptase and integrase, and the plasmid pMDLg / pRRE containing the Rev response element RRE, which was transfected into the human embryonic kidney epithelial cell line 293T for packaging.

[0174] The target plasmids include shRNA of MSX1 and CDX2 and their controls.

[0175] Packaging steps:

[0176] 1) 293T cell passaging requires that the cells are in good condition, with distinct cell boundaries, no clumping or accumulation, and 3-4 synapses. Packaging can be performed when the cells are approximately 80% confluent. This is done using a 10 cm dish as an example.

[0177] 2) Reagent Ratio: The final PEI to plasmid ratio is 1:3 (μL / μg), with 90 μL PEI mixed with 15 μg of target plasmid, 5 μg of pMDLg / pRRE, 5 μg of pRSVREV, and 5 μg of VSV-G.

[0178] 3) Mix the above plasmid with 800 μL Opti-MEM to prepare a plasmid + Opti-MEM solution;

[0179] 4) Mix 90 μL of PEI with 800 μL of Opti-MEM, vortex to mix, and let stand at room temperature for 3 minutes to prepare a PEI + OptiMEM solution;

[0180] 5) Add the PEI + Opti-MEM solution dropwise to the plasmid + Opti-MEM solution, shaking the solution immediately after each drop. Vortex to mix thoroughly and let it stand at room temperature for 15 minutes.

[0181] 6) Add the mixture dropwise to the corresponding 293T cells. Because 293T cells adhere poorly to the cell wall, be gentle. Return the cells to the incubator.

[0182] 7) 12-16 hours after plasmid infection, replace the culture medium with fresh high-glucose DMEM + 10% FBS medium and continue culturing;

[0183] 8) 44-48 hours after transfection, harvest the virus. Remove the transfected 293T supernatant with a syringe, filter, and store at -80°C. Quantitative PCR can be used to measure viral titer.

[0184] To infect hiPSCs with viruses, follow these steps:

[0185] 1) When the hiPSC density reaches about 30%, prepare for infection and remove the virus from -80°C and thaw it in advance;

[0186] 2) Mix PGM1 and virus in a ratio that depends on the titer test results, usually 4:1;

[0187] 3) Add 0.1% Polybrene based on the total volume of the above solution to increase the efficiency of viral infection;

[0188] 4) After the cells are removed, the culture medium is aspirated and the above mixed solution is added;

[0189] 5) 24 hours after infection, the culture medium was replaced with PGM1 and culture continued;

[0190] 6) After another 24 hours, when the viral transgenic gene is fully expressed, puromycin is added for 24 hours to screen for resistance genes;

[0191] 7) The final surviving cells are the cells that are successfully infected by the virus and can continue to proliferate and differentiate.

[0192] 1.7 Immunofluorescence staining

[0193] 1) Fixation: Remove cells, aspirate culture medium, and wash three times with 200 μl / well PBS. Add 200 μl / well of 4% paraformaldehyde (PFA) and fix for 15 minutes at room temperature. Aspirate fixative and wash three times with PBS per well.

[0194] 2) Blocking and Permeabilization: Dilute 2 μl of Triton X-100 with 1 ml of PBS to make a 2% PBST solution. Dissolve 3 μl of donkey serum on ice and dissolve in 1 ml of PBST. Mix well and add to the plate. Block and permeabilize at room temperature for 10 minutes. Blot dry and wash three times with PBS with shaking.

[0195] 3) Thaw the primary antibody on ice, dilute it in 0.1% BSA, and store at 4°C overnight.

[0196] 4) The next day, take out the sample from the refrigerator, remove the primary antibody, and then wash it three times for reuse.

[0197] 5) Melt the primary antibody on ice, dilute it in 0.1% BSA, and incubate it at 37°C in the dark for 1 hour. Remove the plate, blot dry, and wash three times with PBS.

[0198] 6) Add Hoechst (1 μg / ml dissolved in PBS) to cover the cells. Let stand at room temperature for 5 minutes, then aspirate the cells and wash three times with PBS. Observe and photograph under a fluorescence microscope with PBS or store in a dark place.

[0199] 1.8 Purification of DACT-1 cells using the light-activated small molecule

[0200] In this experiment, live cells were incubated with culture medium containing DACT-1 (Halabi et al., 2020). DACT-1 small molecules were activated in regions of interest under 405 nm light. DACT-1 is fixed to the cells by binding to internal proteins in living cells, and its structural changes allow it to emit light under 561 nm laser activation. Therefore, DACT-1 was combined with restricted photoactivation microscopy to label cells in different regions. Purified cells were obtained after flow cytometry sorting.

[0201] Photoactivation experiments were performed on an inverted fluorescence microscope (Nikon TiE) equipped with a motorized stage (Marzhauser SCANIM). The imaging system was equipped with a 20×0.75NA dry objective, a spinning disk confocal unit (Yokogawa CSU-X1), and a scientific CMOS camera (Hamamatsu ORCA-Flash 4.0 v2). The microscope, camera, stage, and lasers were controlled by Micro-Manager (version 2.0.0). Micro-Manager was controlled using an interactive interface in MATLAB R2018b, enabling customizable hardware control (e.g., controlling stage movement along a specific trajectory). Red illumination for DACT-1 confocal imaging was provided by a 561 nm laser (Coherent OBIS 561 nm, 50 mW), and violet activation was provided by a 405 nm laser (Coherent OBIS 405 nm, 50 mW). The specific operational procedures are as follows:

[0202] 1) Dissolve DACT-1 in DMSO to a concentration of 10 mM, aliquot and store at -20°C in the dark.

[0203] 2) Remove the cells at the desired purification stage and replace the original culture medium with RPMI+B27 containing 1 μM DACT-1. Incubate in the dark at 37°C for 30 minutes.

[0204] 3) Start Micro-Manager and control the imaging system through MATLAB, visually inspect and capture DIC images of cells in the 96-well plate under bright field, and transfer the images to MATLAB;

[0205] 4) The region for restricted photoactivation of DACT-1 was selected and drawn as a polygon in MATLAB (R2018b, MathWorks), parallel horizontal traces with a spacing of 20 μm were generated and intersected with the polygon, and the platform coordinates of the intersection points were calculated;

[0206] 5) Focus a 405 nm laser (the laser output power is set to 10% of the total power to reduce light damage) on the sample plane to form a fixed position spot with a diameter of 20 μm;

[0207] 6) MATLAB controls the motorized stage to move along the trajectory at a speed of 0.12 mm / s, achieving relative motion between the cells and the laser. This allows the 405 nm laser to scan the entire selected region and selectively activate cells within the region (RFP). Photoactivation of each ROI is typically completed within 1 minute.

[0208] 7) The effect of photoactivated fluorescent labeling weakens over time, and the remaining labeling rate is less than 50% after 24 hours. Therefore, the flow cytometry step is best performed immediately after photoactivation.

[0209] 8) After restricted photoactivation, cells were dissociated into single cells using the digestion method described above, resuspended in 0.5% BSA, and placed on ice;

[0210] 9) RFP-positive and RFP-negative cells were separated and collected using a BD FACS Aria III cell sorter;

[0211] 10) After counting, resuspend in RPMI+B27 medium and add 1 μM Y27632, then plate in a Matrigel-treated empty plate for subsequent experiments.

[0212] The DACT-1 used in this experiment was provided directly by the laboratory of the author Pablo Rivera-Fuentes.

[0213] 1.9 qPCR fluorescence quantification

[0214] Total RNA was isolated using the Easy Pure RNA Kit, and cDNA was synthesized using the TransScript All-in-One First-Strand cDNA Synthesis SuperMix for qPCR kit. Reverse transcription products were used for quantitative qPCR. Gene expression levels were assessed using GAPDH as an internal reference.

[0215] 1.10 RNA-seq Sample Acquisition and Analysis

[0216] This article involves two sets of RNA-seq results, as follows:

[0217] The first group of samples included 12 samples collected for analysis, including AI-CPCs, non-CPCs, hiPSC-CMs, and hiPSCs (including three biological replicates). AI-CPCs and hiPSC-CMs were purified by DACT-1 photoactivation before collection; non-CPCs were collected on day 6 at a dose that deviated from the appropriate CHIR dose; and hiPSCs were cultured in CDM medium before reaching a differentiation-ready state.

[0218] The second group of samples were all collected in the first stage of differentiation (0-72h), and a total of 10 cell samples with different CHIR doses (hiPSC; CHIR2μM48h, 6μM24h, 6μM36h, 10μM24h, 8μM36h, 6μM48h, 12μM24h, 12μM36h and 10μM48h) were collected.

[0219] use RNA was extracted using the RNAKit (TransGene) and sequenced on a NovaSeq 6000-PE150. Reads were processed and mapped to the human GRCh38 / hg38 genome. Principal component analysis (PCA), heatmaps, and GO analysis bubble plots were performed using the Omicshare tool (https: / / www.omicshare.com / ). For heatmaps, FPKM was first converted to log2(FPKM+1) and normalized across samples. Differentially expressed genes (DEGs) were detected between samples using the DESeq (Huber, 2010) method. Genes with a p-value <0.05, a fold change <0.5 or >2, and an average gene expression >1 were considered DEGs and used for further analysis. DEGs in sample 2 were identified by fold change (FC) within the sample; only genes with a fold change <0.5 or >2 and an average gene expression >1 were considered DEGs.

[0220] 2. Image Acquisition and Analysis Methods

[0221] 2.1 Image Capture and Stitching

[0222] In this experiment, the entire process of stem cell differentiation induction into cardiomyocytes (hiPSC-iCM) takes 10-15 days. The Zeiss Cell Discoverer 7 (CD7) is used to culture and image live cells for a long time. It has a small culture chamber inside that can provide a good culture environment for cells with constant temperature and humidity, and provides CO2 and O2 concentration control modules. In order to complete long-term live cell culture and imaging, the internal culture chamber is set to a constant temperature of 37°C, a constant CO2 of 5% throughout the process, and sufficient water is ensured in the air inlet wet bottle.

[0223] To reduce the phototoxicity of long-term shooting to living cells and improve image clarity, the CD7 is equipped with Hamamatsu's ORCA-Flash4.0V3 lens. Its highly sensitive CMOS (Complementary Metal Oxide Semiconductor) can capture images with higher resolution (2048*2048 pixels) and higher signal-to-noise ratio in a short shutter time.

[0224] After several preliminary experimental attempts, we ultimately settled on a 5X objective lens combined with a 2X magnifier (i.e., 10x magnification). We then binned the images 2x2, improving the signal-to-noise ratio and reducing data storage requirements while preserving the finer details of the cells. The resulting single image was 1024 x 1024 pixels with a resolution of 1.3 μm / pixel.

[0225] According to the above experimental steps, hiPSC-CM differentiation induction is divided into three stages. Manual medium changes are required every 24 or 48 hours: basal medium to ensure normal cell growth and small molecule drugs to ensure transitions between experimental stages. Because each manual medium change requires pausing filming, removing the culture dish from the incubator, and replacing it after medium change, we capture images and save separate files throughout the induction experiment, punctuated by each medium change.

[0226] The culture dishes used in this project are all Falcon brand (the low thickness of the culture dishes provides high uniformity, which facilitates batch repetition of experiments). 24-well, 96-well, and 384-well culture dishes were used in the experiment. The specific shooting settings for the three different well plate sizes are as follows:

[0227] 1) 96-well Plate Image Acquisition: The majority of experiments in this project were cultured and imaged in 96-well plates. Each well of the Falcon 96-well plate has a diameter of 6 mm. To ensure complete image capture, 5x5 tiles were used for capture, and the images were stitched together to create a full-well image. To ensure smooth image stitching, we set a capture coverage of 5%-15%. The 5x5 tiles were stitched together to generate a 4860 x 4860 pixel image (5% overlap) for each well, resulting in a stitched field of view of 6.3 mm x 6.3 mm. To obtain 3D image information of the samples, we captured 3-5 layers of Z-axis images at equal intervals (3-6 μm) along the Z axis. This provides richer sample image information and reserves data for expanded analysis. Based on the above imaging parameters, a single scan of a 96-well plate takes approximately 72 minutes, acquiring a total of 7,200 images (5 rows * 5 columns * 3 layers * 96 wells), and approximately 144,000 images can be acquired in 24 hours. Because the edges of the culture dish are included in the acquired full-well images, which makes image feature analysis difficult, in most of our studies, only nine 3x3 tiles (tiles) located in the center of the culture dish wells were analyzed to avoid interference from the dish edges.

[0228] 2) 24-well plate image acquisition: Similar to 96-well plates, each well in a 24-well plate is composed of 156 tiles stitched together to form a single 20,284 x 20,284 pixel image (10% coverage). Note that because some wells near the edge of the 24-well plate are beyond the microscope objective's range (exceeding the maximum stage movement), only 136 tiles are captured for these wells. Each well has a visible area of ​​approximately 13.0 mm x 13.0 mm, resulting in a total of 10,992 images collected in one cycle (136 tiles x 3 layers x 4 wells + 156 tiles x 3 layers x 20 wells).

[0229] 3) 384-well plate image acquisition: For 384-well plates (square wells), due to the smaller well area, a 3x3 scanning strategy was used, resulting in 9 images (tiles). Using a 10% coverage ratio and capturing only a single tile, a total of 3456 (3 rows * 3 columns * 1 tile * 384 wells) images were acquired in one round of imaging.

[0230] Cell images were captured using Zeiss's ZEN image acquisition software (V2.0-V3.1). Cell images captured by the microscope were saved as raw files in the CZI format. To enable the designed imaging system to perform real-time image processing and decision-making, a script was developed to save the uncompressed images captured in real time as TIFF or PNG formats for easy post-processing.

[0231] 2.2 Image texture feature extraction and manifold analysis

[0232] To analyze texture features in real-time image streams during hiPSC-iCM differentiation induction experiments, we used SIFT, SURF, and ORB feature descriptors to generate 448-dimensional high-dimensional local features. We then used dimensionality reduction methods such as PCA and LDA to analyze and visualize experimental results at different differentiation stages and efficiencies. We implemented this code using the OpenCV and scikit-learn packages in Python.

[0233] 2.3 Assessment of differentiation efficiency using cTnT fluorescence staining

[0234] The differentiation efficiency of iPSCs to CMs in each well was quantified by the average fluorescence intensity of the final fluorescence staining image. Specifically, for a W×W fluorescence staining image I (intensity value ∈ [0, 1]), its “differentiation efficiency index” is defined as the total fluorescence intensity of pixels whose intensity value exceeds the threshold α, that is,

[0235] Where 1 / W 2 is the normalization factor. In our experiments, we take α = 0.2.

[0236] 2.4 Image Analysis of Differentiation hiPSC-CM Stages - GoogLeNet

[0237] 2.4.1 Image Preprocessing

[0238] Due to the different image resolution and quality of different batches, in the experiment, the resolution of bright field and cTNT fluorescence images of the hiPSC-CM stage was first adjusted to 2816 × 2816 pixels, and the contrast and brightness of the fluorescence images were adjusted.

[0239] Specifically, to enhance contrast, the fluorescence images were processed using a contrast-limited adaptive histogram equalization algorithm (Zuiderveld, 1994) or a low-light image enhancement algorithm (Xuan et al., 2011) to achieve comparable contrast. Regarding brightness, these fluorescence images were converted to the HSB (hue-saturation-brightness) color space, and the brightness values ​​were multiplied by 0.8.

[0240] In this stage of the image processing framework, the brightfield image is sliced ​​into tiles, and image classification and transformation are performed on a tile-by-tile basis. To obtain these tiles, both the full brightfield and fluorescence images are cropped into 512×512 tiles, with a 50% overlap between adjacent tiles; thus, the entire image is sliced ​​into exactly 100 tiles. All of the above image preprocessing steps were implemented using MATLAB (R2020a, MathWorks).

[0241] 2.4.2 Tile Classification

[0242] To classify image patches as "0" (negative, meaning almost no hiPSC-CMs) or "1" (positive, meaning typical hiPSC-CMs), this experiment used the classic deep convolutional neural network classifier, GoogLeNet (Szegedy et al., 2015). A dataset of n = 1354 brightfield cell patches was constructed and labeled as "0" or "1" based on the final cTNT fluorescence results. These patches were then randomly divided into a training set (n = 945) and a test set (n = 409). 30% of the patches in the training set were used for validation. During training, RMSprop (Hinton et al., 2012) was used as the optimizer, with a mini-batch size of 66, a learning rate of 0.0001, and an l2 regularization parameter of 0.0001. GoogLeNet was trained for 10 epochs. GoogLeNet was implemented using MATLAB (R2020a, MathWorks) and trained on a GPU with 8GB of video memory.

[0243] 2.4.3 Image Conversion from Brightfield to Fluorescence

[0244] In order to convert the bright field tiles of the hiPSC-CM stage into fluorescent tiles, for the bright field tiles predicted by GoogLeNet as negative and positive, this experiment used two independent CycleGANs (Zhu et al., 2017) to convert the two types of tiles into cTNT fluorescence images (Figure 7). CycleGAN is one of the most popular deep generation models for image conversion. Here, x∈X represents the bright field tiles of the hiPSC-CM stage, and y∈Y represents its corresponding fluorescent tiles (where X and Y are the datasets of bright field and fluorescent tiles, respectively). In the CycleGAN model, the generator G X→Y Perform the conversion from bright field image to fluorescence image, and introduce the additional generator G Y→X Perform the conversion in the opposite direction. During training, introduce the adversarial loss L (advY) ,L (advX) (defined as the discriminator D Y and D X The best classification performance of G is obtained by using these two discriminators to distinguish between real and generated patches) X→Y and G Y→X , which is to encourage them to generate patches that are very similar to the patches in the dataset; introducing the cycle consistency loss L cyc (weight is λ) to further encourage G Y→X (G X→Y (Y))≈x and G X→Y (G Y→X (y))≈y. Since CycleGAN was originally designed for unpaired image conversion, and the dataset used in this experiment provides paired bright field and fluorescence patches, an additional loss term is added

[0245] (weight is μ, where HW finds the total number of pixels of a patch) to explicitly induce the generated fluorescence patch G X→Y (x) is similar to the actual fluorescence result y. Therefore, the total loss function is modified to

[0246] L=L adv(Y) +L adv(X) +λL cyc +μL sim .

[0247] In this experiment, a dataset consisting of 3500 pairs of brightfield images of hiPSC-CM stages and corresponding cTNT fluorescence images was constructed for training, and 3600 pairs for testing (from 35 and 36 pairs of complete images). Based on the predictions of the trained GoogLeNet (Tables 3.1 and 3.2), the dataset was divided into negative and positive datasets for training and testing CycleGAN-0 and CycleGAN-1, respectively (Figure 8). Both CycleGANs were trained using the Adam optimizer (Kingma and Ba, 2015) with parameters β1 = 0.5 and β2 = 0.999. The initial learning rate was set to 0.0002, and the learning rate schedule was consistent with (Zhu et al., 2017). Both regularization parameters were set to λ = 4 and μ = 10. CycleGAN-0 and CycleGAN-1 were trained for 50 and 100 epochs, respectively. Finally, when predicting the cTNT fluorescence image for the entire brightfield image on the test set, the output tiles generated by the two trained CycleGANs were reconstructed to obtain the complete fluorescence image prediction; during concatenation, the prediction value for the overlapping area was taken as the average of the overlapping tiles. CycleGAN was implemented using the PyTorch framework (Paszke et al., 2019) and trained on a GPU with 8GB of video memory.

[0248] 2.4.4 Evaluation of Model Performance

[0249] The classification performance of GoogLeNet is evaluated by accuracy (ACC), precision, and recall (also known as sensitivity). They are defined as follows:

[0250] Among them, "#" represents the "number of...", "TN", "TP", "FN", and "FP" represent "true negative", "true positive", "false negative", and "false positive", respectively. They all range from 0 to 1, and the higher the value, the better the classification performance. In order to evaluate the performance of bright field to fluorescence image conversion, it is noted that the hiPSC-CM differentiation efficiency of the hole can be quantified by the total fluorescence intensity of the cTNT fluorescence image (named "differentiation index"), so here we consider measuring the model performance by the consistency of the differentiation index of the real and predicted cTNT fluorescence images. Specifically, for the fluorescence image I (grayscale value ∈ [0,1]), its "differentiation index (Differentiation Index)" is defined as the sum of the grayscale values ​​greater than the threshold α and then normalized, that is,

[0251] In this experiment, M = N = 2816 (after preprocessing), and the threshold α was chosen to be 0.15. The performance of image conversion was measured by the Pearson correlation between the differentiation index of the real and predicted fluorescence images in the test set (n = 36); the high correlation indicates that our method can accurately predict the final differentiation efficiency from bright-field images of cells at the hiPSC-CM stage.

[0252] 2.5 Image Analysis of Cardiac Differentiation hiPSC-CM Stages - pix2pix Model

[0253] 2.5.1 Machine Learning Model

[0254] We chose the pix2pix model (Isola et al. 2017) to predict fluorescence images based on brightfield images. In the pix2pix model, the generator G generates corresponding fluorescence images based on the brightfield image, while the discriminator D learns to distinguish between real and fake "brightfield-fluorescence" image pairs (as shown in Figure 38a). Formally, let x and y denote the brightfield image and the corresponding fluorescence image, respectively, and z denote the randomness in the generator G. The final training objective is the L1 reconstruction loss (weighted by λ) plus the adversarial loss, i.e.,

[0255] in

[0256] Following the design of (Isola et al. 2017), the generator G is based on the classic U-Net architecture (Ronneberger, Fischer, and Brox 2015). The transposed convolution module is replaced with nearest neighbor upsampling followed by ordinary convolution to avoid checkerboard artifacts (Odena, Dumoulin, and Olah 2016). We use instance normalization (see Figure 38b). The discriminator D is a patch discriminator that outputs a classification score map with a receptive field of 16×16 pixels per pixel in the original image (see Figure 38c).

[0257] 2.5.2 Experimental Setup and Evaluation

[0258] All images were resized to 1,536 × 1,536 pixels. During each training epoch, 1,260 patches of 256 × 256 were randomly cut from the training set images. The batch size was 16. The model was trained for 2,000 epochs using the Adam optimizer (Diederik P Kingma and Ba 2015) with parameters β1 = 0.5 and β2 = 0.999. λ was set to 100. The learning rate was fixed at 0.0002 in the first 1,000 epochs and then linearly decayed to 0 in the next 1,000 epochs. To further ensure the fidelity of the fluorescence predictions, the adversarial loss was disabled for the last 1,000 epochs.

[0259] During testing, the input is the entire image. We compare the predicted fluorescence map with the true fluorescence map pixel by pixel, displaying the results as a heatmap and calculating the Pearson correlation coefficient. We also compare the predicted and true differentiation indices at the full image scale.

[0260] 2.6 Image Analysis of Cardiac Differentiation hiPSC-CPC Stages

[0261] 2.6.1 Image Annotation and Image Preprocessing

[0262] Image annotation and preprocessing were implemented in MATLAB (R2018b, MathWorks). To complete the annotation of training and test image masks, this paper tracked live cell bright field images from day 6 to the end of differentiation. Specifically, this paper tracked the cTNT area in the image stream from day 6 to day 12 of differentiation. Further, based on the experience of experts, the CPC area was manually annotated on the bright field image on day 6 to obtain the corresponding mask. The annotated bright field image mask contains dark gray, light gray, and black areas: cell areas with a high probability of successful differentiation into hiPSC-CMs and typical texture are marked in dark gray; cell areas with difficult to predict whether they can be successfully differentiated from the texture, or cell areas at the edge of successfully differentiated cells tracked by image stream are marked in light gray; the remaining cell areas with little possibility of differentiation into hiPSC-CMs are marked in black.

[0263] During the preprocessing process, all batches of images in this experiment (including live cell bright field images on day 6, manually annotated masks, and cTNT immunofluorescence images) were uniformly resized to 2816×2816 pixels. In the subsequent weakly supervised learning process, the resized complete images were further divided into tiles (512×512 pixels). During the cutting process, there was a 50% (75%) overlap between adjacent tiles in the training and validation sets (test sets). Therefore, each complete image was divided into 100 (361) tiles in the training and validation sets (test sets). The preprocessed dataset contains multiple groups of images from different batches, and the detailed information is shown in Table 5.

[0264] 2.6.2 Weakly Supervised Learning

[0265] In this experiment, the ResNeSt-101 (Zhang et al., 2020a) network was used to determine whether there were CPC areas that could differentiate into cardiomyocytes in the bright field images on the 6th day (Figure 39a). The label of each bright field tile was divided into credible labels and uncertain labels based on the corresponding manually annotated mask tile. Specifically, if the dark gray area of ​​the mask tile accounts for more than 30% or the entire tile is black, the bright field tile label corresponding to the mask tile is defined as a credible label "1" or "0" respectively; and the labels of the remaining bright field tiles are all considered uncertain labels. Only tiles with credible labels were used for training and validation of the weakly supervised learning model, while all types of tiles were used for testing the model. The Adam optimizer was used in the training process, and the loss function was the cross-entropy loss function. The trained model was used to perform binary classification on the bright field tiles in the test set.

[0266] The classification results include 0 and 1, where 0 indicates that the model predicts that the bright field patch does not contain CPC regions that can differentiate into hiPSC-CMs. Conversely, 1 indicates that the model predicts that the bright field patch contains CPC regions that can differentiate into hiPSC-CMs. Furthermore, this experiment used Grad-CAM (Selvarajue et al., 2017) to locate CPC regions that can differentiate into hiPSC-CMs in the bright field patch (Figure 39b). Specifically, Grad-CAM combines the final convolutional layer of ResNeSt-101 and the back-propagation gradient of the specified target category (label 1) flowing through the final convolutional layer to generate saliency patches and binary patch results corresponding to the bright field patch (Figure 15).

[0267] The highlighted areas in the saliency patches are the basis for ResNeSt-101 predicting the label of the brightfield patches as 1, indicating that these areas contain CPC textures that can successfully differentiate into hiPSC-CMs. For brightfield patches classified as 0 by the model, their binarized patches are directly set to black; for brightfield patches classified as 1, the corresponding saliency patches are binarized using a threshold of 10 (pixel values ​​greater than 10 are set to 255, white; otherwise, they are set to 0, black).

[0268] 2.6.3 Model Performance Evaluation

[0269] This paper evaluates the performance of weakly supervised learning models from three different perspectives: neural network classification performance, prediction metrics calculated based on manually annotated masks, and prediction metrics calculated based on cTNT immunofluorescence images. The specific methods are as follows:

[0270] 1) Neural network classification performance

[0271] The classification performance of ResNeSt-101 used in the weakly supervised learning model in this paper is evaluated by accuracy (ACC) and area under the curve (AUC).

[0272] 2) Prediction indicators calculated based on manually annotated masks

[0273] The binarized patches generated by Grad-CAM are used for comparison with manually annotated masks. Before calculating the indicators, the binarized patches need to be reconstructed into complete images. The reconstruction principle is that the overlapping parts between patches with different prediction results are preferentially set to white (CPC areas that can differentiate into hiPSC-CMs). In order to evaluate the pixel-level classification performance of the model, this paper calculates a series of prediction indicators, including accuracy, F1 coefficient, precision, recall, specificity and intersection over union (IoU). They are defined as follows

[0274] The "#" represents the number of pixels, and "TN," "TP," "FN," and "FP" represent true negatives, true positives, false negatives, and false positives, respectively. They all range from 0 to 1, with higher values ​​indicating better performance.

[0275] During the calculation process, both dark gray and light gray areas in the manually annotated mask were considered as CPC regions that could differentiate into hiPSCCMs and were used to match the white areas in the binary image.

[0276] 3) Prediction indicators calculated based on cTNT immunofluorescence images

[0277] In this paper, the Pearson correlation coefficient was used to evaluate the match between the predicted and actual differentiation efficiencies. The predicted differentiation efficiency was simply defined as the proportion of white area in the reconstructed binary image, while the differentiation efficiency index defined above was used to measure the actual differentiation efficiency in the cTNT immunofluorescence images.

[0278] The Pearson correlation coefficient was calculated based on the predicted differentiation efficiency from the day 6 brightfield images and the corresponding differentiation efficiency index from the cTNT immunofluorescence images. The correlation coefficient fell within the interval [0, 1], providing an approximate assessment of the reliability of the predicted cell differentiation results. Because cTNT immunofluorescence images can vary between batches, the correlation indices were calculated for the same batch to ensure comparability of the results.

[0279] 2.7 Image Analysis of Cardiac Differentiation into Mesoderm

[0280] 2.6.1 Preparation of labeled dataset

[0281] The training and validation of the classification system described in this article requires a dataset consisting of bright field image streams of each well, and each well is labeled ("low", "moderate" or "high"). Because hiPSCs do not respond uniformly to CHIR, moderate CHIR experimental conditions may vary from batch to batch. Therefore, for each specific CHIR duration condition (24 hours, 36 hours, or 48 hours), if the average percentage of cTNT-positive cells in the wells using a certain CHIR concentration is ≥20%, then the CHIR concentration is first determined to be a "moderate" concentration; CHIR concentrations outside the moderate concentration range are labeled as "low" or "high". For each concentration level c, its relative difference from the moderate concentration range [c1, c2] is defined as "ΔCHIR concentration", that is

[0282] This is used to measure the degree of deviation of the concentration from the neutral condition. Following the above steps, each well can be assigned a label for each CHIR duration condition. The four batches of labels used in this phase of the experiment are listed here for CHIR durations of 24, 36, and 48 hours (Table 6).

[0283] 2.7.2 Cell Image Preprocessing

[0284] The image resolution, brightness, and contrast of each well in the dataset may vary. To obtain a uniform feature representation, the resolution, brightness, and contrast of the input live cell brightfield image stream were normalized. First, all images were resized to 4860 × 4860 pixels with a grayscale value range of 0 to 255. Second, the image stream of each well was gamma-corrected, so that the median grayscale value was transformed to approximately 127. Finally, grayscale values ​​below and above the median were processed twice using gamma transformations, respectively, so that the lower and upper quartiles of the grayscale distribution were transformed to approximately 96 and 160.

[0285] 2.7.3 Image Stream Feature Extraction

[0286] The image stream for each well consists of 10 brightfield images (at timestamps T1, T2, …, T10), taken at equal intervals between 0 and 12 hours during the first stage of differentiation. Several image features potentially relevant to the classification task were designed, including fractal dimension, cell coverage statistics (area, perimeter, area-perimeter ratio, brightness, local entropy), and optical flow (texture features were also explored but appear irrelevant to classification; data are not presented here). Of these features, "optical flow" was calculated for every two consecutive timestamps (this type of feature is designated as Type-II features), while the others were calculated for each timestamp (this type of feature is designated as Type-I features) (Figure 3.26c); in both cases, a real number sequence was generated to represent the feature value. The values ​​of "area," "perimeter," "area-perimeter ratio" ("AC ratio"), and "optical flow" were then normalized by dividing them by the first value in the sequence (referred to as "relative features"); the other features were used without normalization (referred to as "absolute features"). Finally, timestamps T1-T10 are divided into prophase, metaphase, and anaphase, and the average of the features for each stage is calculated (Figure 26c). Therefore, each of these seven features will give three real numbers (corresponding to prophase, metaphase, and anaphase), resulting in a 21-dimensional feature representation for each well.

[0287] The calculation details of each feature are listed below. They are calculated using the Python package scikit-image (Van Der Walt et al., 2014).

[0288] 1) Fractal dimension. The fractal dimension measures the roughness and self-similarity of an image. This experiment uses the differential box counting method (Sarkar and Chaudhuri, 1994) to calculate the fractal dimension of an image (in the range of 2 to 3). The box widths are selected as 2, 2k, 2k2, …, 2k 15 ; k is chosen as (243) 1 / 15, so that the width ranges from 2 to 243 (1 / 20 of the image width).

[0289] 2) Local entropy. For each pixel in a given image, the entropy of the grayscale distribution (ranging from 0 to 255) of pixels with a Euclidean distance of ≤10 (pixels) to it is calculated. Because the local entropy of cell-free regions is low, we simply set the threshold to 3 and discard pixels with a local entropy <3. The average local entropy is then used as the final result.

[0290] 3) Area, perimeter, and area-to-perimeter ratio. Similar to (2), pixels with a local entropy ≥ 3 are considered to be covered by cell clones. The area is then the number of pixels covered by the cell clone, and the perimeter is the total length of the cell clone outline. The area-to-perimeter ratio (AC ratio) is the area divided by the perimeter and reflects the proportion of cells located at the edge of the clone.

[0291] 4) Cell brightness. Again, the local entropy criterion is used here to detect regions with cell clones. Therefore, cell brightness is their average grayscale value, which may be related to the compactness of the cells.

[0292] 5) Optical flow. Optical flow is a common method used in image flow analysis to estimate the motion of objects between consecutive frames. Here, it can be used to measure cell movement during differentiation, which reflects the shrinkage rate of cell clones. Here, the algorithm of GunnerFarneback ( 2003) to estimate dense optical flow between two consecutive time-stamped images, with the following parameters: pyramid scale = 0.5, pyramid levels = 3, window size = 16, number of iterations = 3, poly_n = 5, and poly_sigma = 1.2. Finally, the average modulus of the optical flow vectors is calculated as the optical flow eigenvalue. Flow vectors with a modulus ≤ 4 are discarded, as these insignificant motions may be due to noise.

[0293] 2.7.4 Feature Space Visualization

[0294] This experiment uses linear discriminant analysis (LDA) (Hastiee et al., 2009) and t-SNE (Van Der Maaten and Hinton, 2008) to visualize the high-dimensional feature space (21 dimensions if all features are used, 4 dimensions if feature selection is performed). LDA (Hastiee et al., 2009) is a supervised dimensionality reduction method that linearly projects the feature space into a subspace with the highest discriminative power. Therefore, LDA can be used to visually examine the discriminative power of feature representations. t-SNE (Van Der Maaten and Hinton, 2008) is an unsupervised nonlinear dimensionality reduction method that also transforms feature representations into lower-dimensional representations, but its dimensionality reduction goal is to preserve the original distance distribution between neighbors as much as possible. Therefore, t-SNE is more suitable for directly visualizing feature distributions. The Python scikit-learn (Pedregosa et al., 2011) package was used to implement LDA and t-SNE. For LDA, the parameter "shrinkage" (l2-regularization coefficient) was set to 0.1 and 0, respectively, when visualizing the 21-dimensional and 4-dimensional feature spaces under a CHIR duration of 24 hours (Figure 27b). For t-SNE, the parameter "perplexity" was set to 130 for visualizing the 21-dimensional feature space under all CHIR duration conditions; the parameter "perplexity" was set to 130, 300, and 200 for visualizing the 4-dimensional feature space under CHIR duration conditions of 24 hours, 36 hours, and 48 hours, respectively, to obtain better visualization (Figures 27a, c, d).

[0295] In addition, we can use the dimensionality reduction techniques LDA and PCA to visualize the high-dimensional feature vectors (21 dimensions if all features are used, 4 dimensions if only selected features are used). LDA is used to verify the discriminative power of feature representation, and PCA is used to visualize the sample distribution. When visualizing the 21-D and 4-D feature spaces, the shrinkage parameter of LDA is set to 0.1 and 0, respectively.

[0296] 2.7.5 Logistic Regression

[0297] Logistic regression is a linear model for classification (Hastie et al. 2009). The training data was reweighted to account for class imbalance. When all 21 features were used, l1 regularization with coefficients of 1 / 4, 1 / 8, and 1 / 8 was used for models with CHIR durations of 24, 36, and 48 hours, respectively, to encourage sparse parameters; when only four selected features were used, l2 regularization with a coefficient of 0.1 was used. The final loss function was optimized using the liblinear solver. Accuracy, precision, recall, F1 score, and AUC were used to evaluate the performance of logistic regression. Precision, recall, F1 score, and AUC were averaged across the three categories.

[0298] By averaging the predictions for wells with concentration c, the logistic regression model can also provide a "bias score" for concentration level c. Let N c is the number of wells with concentration c, where Wells are predicted as low, optimal, and high by logistic regression. Then, the deviation score is defined as:

[0299] The deviation score ranges from -1 to 1, which can reflect the deviation of CHIR concentration from the optimal condition.

[0300] 2.7.6 Cross-batch Validation

[0301] To test the model's generalization to new batches, cross-batch validation was performed with a CHIR duration of 24 hours. Feature selection was performed to improve classification generalization. In each training-test cycle, one batch was selected for testing, while the remaining batches were used for feature selection and training. The logistic regression model in each cycle was regularized using elastic-net (with an l_1 fraction of 0.1 and a weight of 0.05) and optimized using the SAGA solver. Cross-batch validation was evaluated using Person correlation sparsity between the predicted deviation scores and the true "ΔCHIR concentration."

[0302] 2.8 Initial iPSC clone status control

[0303] We prepared a dataset of n = 1934 whole-well bright field images of initial iPSC colonies at 0 h (before CHIR treatment). 343 features were extracted from the bright field images to quantify the morphological characteristics of the initial iPSC colonies, as follows:

[0304] (9) Local entropy, cell brightness, cell contrast, and total variation. Local entropy is the average local entropy of each pixel in the cell-containing region, where the local entropy of a pixel is calculated from the intensity distribution of its neighborhood with a radius of 5 pixels. Cell brightness and cell contrast are the mean and standard deviation of the intensity of the cell-containing region. Total variation is the L1 norm of the brightfield image gradient.

[0305] (10) Hu invariant moments 1 to 7. These are the seven image moments of bright field images, which are invariant to translation, scaling, and orthogonal transformations.

[0306] (11) SIFT 1-256. These are 256-dimensional “bag of keypoints” representations using the SIFT feature descriptor. Specifically, we first apply K-Means on the SIFT feature vectors of all keypoints in 385 bright field images (not included in the dataset) to obtain 256 classes; then, for each image in the dataset, we count the number of keypoints assigned to each class, resulting in a 256-dimensional feature vector.

[0307] (12) ORB 1 to 64. Similar to SIFT 1 to 256, ORB 1 to 64 is a 64-dimensional “bag of keypoints” representation using the ORB feature descriptor.

[0308] (13) Area, perimeter, and area / perimeter ratio. These are the total area, total perimeter, and their ratio of the cell-containing region. The area and perimeter are normalized by the square of the image width and the width, respectively.

[0309] (14) Solidity, convexity, and roundness. For a connected component R, solidity is defined as Convexity is defined as Roundness is defined as For bright field images, its solidity, convexity, and circularity are the average of the solidity, convexity, and circularity of the connected components of all cell regions, respectively, and are weighted by the area of ​​the connected components.

[0310] (15) Maximum center-contour distance (CCD), minimum CCD, minimum / maximum CCD ratio, mean CCD, and standard deviation of CCD. For each connected component, we calculated the distribution of distances between the center and the boundary points. Statistics (minimum, maximum, minimum / maximum ratio, mean, and standard deviation) were calculated from the distribution. For the entire bright field image, these features were also weighted averages of the values ​​of all connected components of the cell area.

[0311] (16) Spacing. To measure the spacing between cell regions, the cell-free region was skeletonized, and the average distance between the skeleton and the cell region was calculated as the spacing.

[0312] We collected cTnT fluorescence images on day 12 to determine the optimal CHIR conditions for each batch. Because the differentiation potential of different cell lines varies even under optimal CHIR conditions, the differentiation efficiency index of each well was normalized to the maximum differentiation efficiency index of that cell line.

[0313] We developed a random forest regression model to predict the final differentiation efficiency index from 343 features of the initial iPSC clones. 1,350 wells were used for training and 584 wells for testing. To determine feature importance, the random forest model employed 1,000 decision trees with a maximum depth of 8. Fifteen features were considered at each branch of the decision tree. For efficiency prediction, the number of decision trees in the random forest model was set to 20.

[0314] 2.9 Early Concentration Assessment of Renal Differentiation

[0315] 2.9.1 Experimental Preparation

[0316] iPSCs and ESCs were resuspended in PGM1 medium (CELLAPY) and seeded with 10 μM Y27632 (Selleck Chemicals) in 24-well Matrigel-coated (Corning) plates. From day 0, the culture medium was changed to Advanced RPMI-1640 (Gibco) with the addition of 1% penicillin-streptomycin (Life Technologies) and 1% GlutaMAX supplement (Gibco). 2-15 μM CHIR (Selleck Chemicals) was added to the culture medium for 4 days (days 0-4), followed by treatment with 10 ng / mL Activin A for 3 days (days 5-7), and then with 10 ng / mL FGF9 for 2 days (days 8-9). On day 9, cells were collected for immunofluorescence staining of SIX2. Wherein "#" represents the "number of..." Therefore, the deviation score ranges from -1 (when all wells are predicted to be "low") to 1 (when all wells are predicted to be "high"), which can indicate to the experimenter the direction in which the CHIR concentration condition deviates from the moderate concentration condition.

[0317] 2.9.2 Dataset

[0318] We prepared a dataset of day 4 bright-field images of renal progenitor cells and assigned concentration labels ("low," "optimal," and "high") to each image based on CHIR dose conditions and final immunofluorescence results. The dataset was randomly divided into a training set (n = 3,398) and a test set (n = 1,457). We used 256-dimensional "bag of keypoints" feature vectors derived from the SIFT feature descriptor as local features for the bright-field images. T-SNE was used for feature visualization, with a perplexity of 60.

[0319] 2.9.3 Logistic Regression

[0320] Logistic regression was used to classify bright-field images into "low," "optimal," and "high" CHIR dose groups. The training data was reweighted to account for class imbalance. The logistic regression model was trained with L1 regularization and optimized using the liblinear solver. Accuracy, precision, recall, F1 score, and area under the curve (AUC) were used to evaluate the performance of the logistic regression. These values ​​were averaged across the three classes.

[0321] 2.10 Image Analysis of Hepatic Differentiation to Endoderm Stage

[0322] The differentiation of hepatic differentiated endoderm (DE) cells was based on a protocol for the induction of hepatocyte-like cells using small molecule compounds. Briefly, iPS-B1, iPS-18, and iPS-M were seeded in 24-well plates and cultured in PGM1 medium. When iPSCs reached the desired confluence, the culture medium was changed to RPMI+B27- medium supplemented with CHIR and IDE1 (MedChem Express). After 24 hours, the culture medium was changed to RPMI+B27- medium containing the previous concentration of IDE1 for 2 days. In order to obtain results with different efficiencies, the iPSC confluence, CHIR concentration, and IDE1 concentration were fine-tuned in several wells according to the experimental design. The culture medium was changed every day. On day 3 (DE stage), the cells were fixed for immunofluorescence staining of SOX17. Live cell bright field images and SOX17 fluorescence images were captured.

[0323] We applied a weakly supervised learning model to identify endoderm cell regions. Because SOX17 is localized in the nucleus, SOX17+ cell regions were identified by performing a morphological closing operation on the binarized SOX17 fluorescence images. The training dataset consisted of eight whole-well bright-field images (resized to 16,000 × 16,000 pixels), which were cropped into tiles (512 × 512 pixels) with a 25% overlap between adjacent tiles. Based on the SOX17 fluorescence results, these tiles were labeled "positive" (≥20% SOX17+ cell region), "negative" (no SOX17+ cell region), or excluded from the training set.

[0324] After training for 300 epochs, the model was tested on 45 new brightfield images (5120 × 5120 pixels in size), which were cropped into tiles (512 × 512 pixels) with a 50% overlap between adjacent patches. The prediction results (Grad-CAM heatmaps) for each brightfield image were reconstructed from the tile-level results.

[0325] Example 1: Image acquisition and overall analysis of differentiation system

[0326] 1.1 Establishment of differentiation system

[0327] This article established a monolayer cardiomyocyte differentiation system (Figure 1) based on a previously reported and widely used cardiomyocyte differentiation method (Aguilar et al., 2015). Human hiPSC monolayers were cultured and differentiation began when confluence reached approximately 80%. In the first stage (0-72 hours), the WNT signaling pathway activator CHIR99021 (CHIR) was used; in the second stage, the WNT signaling pathway inhibitor IWR1 was treated for 48 hours; in the third stage, insulin was added to the basal differentiation medium, allowing the cells to spontaneously differentiate into beating cardiomyocytes. The entire process progressed through four stages: stem cells (hiPSCs), cardiac mesoderm (Stage I), cardiac progenitor cells (CPCs, Stage II), and cardiomyocytes (CMs, Stage III). Typically, beating cardiomyocytes can be observed under a microscope within 7-10 days.

[0328] 1.2 Identification of differentiated cardiomyocytes

[0329] After establishing the cardiomyocyte differentiation system, hiPSC-CMs were characterized. Immunofluorescence staining revealed the expression of cardiomyocyte-specific proteins such as cTNT, GATA4, NKX2.5, MEF2C, and α-ACTININ (Figures 2a, b). α-ACTININ staining revealed clear sarcomere structure under a standard fluorescence microscope (Figure 2b). qPCR analysis revealed significant upregulation of cardiomyocyte-specific genes, including those associated with cardiac sarcomeres, various ion channel genes, and metabolism-related genes. However, the maturity of the differentiated hiPSC-CMs remained inferior to that of primary cardiomyocytes (Figure 2d). Patch-clamp electrophysiology revealed that the majority of hiPSC-CMs exhibited action potential behavior consistent with ventricular myocytes, with a plateau phase. A minority of cells displayed characteristics of atrial myocytes, with relatively stable action potentials during measurement, but elevated resting potentials. Furthermore, the cell beating frequency was unstable, and the calcium flux was weak, indicating suboptimal cardiomyocyte maturity (Figure 2c), consistent with previously reported findings in hiPSC-CMs.

[0330] 1.3 Differentiation system is unstable

[0331] During the establishment of the myocardial system, we found that even if the conditions and operations were kept constant as much as possible between batches, the differentiation efficiency was still unstable. When using different stem cell lines for differentiation, the optimal CHIR conditions varied significantly. For example, for different cell lines, if the CHIR treatment time in the first stage of fixed differentiation is 24 hours, the optimal CHIR small molecule concentration required for different cell lines is different, and the CHIR concentration range applicable to each cell line varies (Figure 3a). Even if the same cell line is used, the same starting cell density is maintained, the same CHIR treatment time and concentration are used, and the operation of the experimental operator is unified throughout the process, the differentiation efficiency between different batches still fluctuates greatly. Some batches may fail to differentiate completely (Figure 3b).

[0332] 1.4 Acquisition of images throughout the myocardial differentiation process

[0333] The CD7 live cell imaging platform equipped in the laboratory enables long-term cell culture and imaging. In this experiment, iPS18, iPS-B1, and H9 cell lines were imaged throughout the differentiation process in 24- or 96-well plates. Brightfield images were acquired using a 10x microscope, with each field of view captured approximately every 72 minutes. HiPSC-CM differentiation was recorded continuously for 10-15 days. Using the open API of ZEN software, the cell culture image stream can be exported in real time. Because all wells of the culture dish must be imaged, and the microscope's field of view cannot encompass the entire well at once, small tiles (tiles) are captured and scanned before being stitched together to create a full-well image. After preprocessing the acquired images, including enhancement and compression, the tiles are cropped and stitched based on their relative positions, resulting in the resulting image stream shown in Figure 4. After differentiation, the cells were immunofluorescently stained (cTNT), and fluorescent images were captured again from the same field of view to document the differentiation results. In addition, to avoid focus drift caused by the uneven bottom of a plastic culture dish when shooting a large area on the dish, we added scanning on the Z axis, which can provide three-dimensional information for subsequent analysis.

[0334] Because differentiation is inherently unstable and to ensure successful differentiation images are captured every time, we introduced several variables during the differentiation process, including cell line type, starting cell density, and CHIR treatment time and concentration. These variables can significantly impact the outcome of cardiomyocyte differentiation. Even small changes in CHIR treatment time and concentration can significantly impact the results, and the optimal CHIR treatment time and concentration vary between batches. Based on this, we acquired a series of brightfield image streams of live cells demonstrating both successful and failed differentiation.

[0335] 1.5 Mapping the cardiomyocyte differentiation trajectory

[0336] We next investigated whether brightfield images contain sufficient features to indicate differentiation status. By extracting 448 dimensions of local features (SIFT, SURF, and ORB) from whole-well brightfield images of iPSC-to-CM differentiation, we found that the distribution of local image features differed between cells with high and low differentiation efficiencies and at different differentiation stages, as shown in principal component analysis (PCA) score plots (Figures 5a, b). Linear discriminant analysis (LDA)42 showed that image trajectories gradually differentiated over time at different CHIR doses (Figures 5c, d). These findings suggest that brightfield image streams contain cues reflecting iPSC differentiation stage, differentiation efficiency, and CHIR dose.

[0337] Example 2: Differentiation efficiency evaluation based on bright field images of hiPSC-CM stage

[0338] 2.1 Deep Learning Method Based on HiPSC-CM Brightfield Images - GoogLeNet

[0339] Next, we examined the local features of various cell lines in brightfield images in more detail. After carefully examining live cell brightfield images of each well, we noted that successfully differentiated cTnT+ CMs had characteristic features: a more compact, domed, three-dimensional morphology; furthermore, successfully differentiated CMs were often connected into sheets or ropes. The morphology of cells that did not differentiate into CMs was heterogeneous, with no obvious aggregation pattern (Figure 6a,b). These findings support the feasibility of using brightfield images alone to identify CMs.

[0340] We use deep learning methods to formulate the problem of evaluating myocardial differentiation as a prediction problem based on brightfield images versus immunofluorescence images. We design a deep learning framework consisting of a patch classification model, GoogLeNet (Szegedy et al., 2015), and two patch conversion models, CycleGANs (Zhu et al., 2017), that convert brightfield images to fluorescence images (Figure 7).

[0341] For the specific image learning process, the full-sized brightfield image was first cut into patches. To eliminate edge effects in the prediction results, the patches had a certain degree of overlap. To train GoogLeNet, the patches were divided into two categories: "0" (negative, i.e., brightfield patches containing almost no hiPSC-CMs) and "1" (positive, i.e., patches containing typical hiPSC-CMs). A dataset of patches labeled "0" and "1" was constructed and randomly used for training and testing (n = 945 training patches, n = 409 test patches) (Figure 8, Table 1).

[0342] Table 1. Summary of image data used by GoogLeNet

[0343] The experiment classified patches into "0" and "1" categories and evaluated GoogLeNet's patch-level classification performance on the training set (n=945) and test set (n=409) using accuracy, precision, and recall. The trained GoogLeNet performed exceptionally well, achieving 94.38% accuracy and 94.55% precision on the test set (Table 2). This demonstrates that the basic classification of patches as positive or negative is accurate.

[0344] Table 2. Classification performance of the tile classification module (GoogLeNet)

[0345] To convert the complete brightfield image into a fluorescent image, we used the previously trained GoogLeNet to classify the corresponding patches into either "0" or "1" categories. These patches were then paired with the corresponding fluorescent patches and fed into CycleGAN-0 and CycleGAN-1 networks for learning (Figures 7 and 8). Specifically, to train the image conversion model, CycleGANs, we paired the brightfield patches with the corresponding fluorescent patches. We then constructed two datasets (n=2057 training and n=2022 test) for training and testing CycleGAN-0 and CycleGAN-1, respectively (n=1443 training and n=1578 test) (Table 3). Given the diverse morphology of differentiated cardiomyocytes, hiPSC-CM images containing a variety of brightfield morphologies were intentionally selected for training and prediction (Figure 6).

[0346] Table 3. Summary of image data used by CycleGANs

[0347] Accurately assess differentiation efficiency using hiPSC-CM images

[0348] After training, CycleGANs performed well on the test dataset, with real cTNT immunofluorescence images and predicted cTNT fluorescence images showing a high degree of similarity (Figure 9a, b, c). Analysis of the results shows that some non-cardiomyocytes with similar morphology to cardiomyocytes or inaccurately focused brightfield images can introduce certain errors in the prediction results.

[0349] For quantitative analysis, we used the Differentiation Index to score cTNT fluorescence images. The Pearson correlation coefficient (r) between the true cTNT immunofluorescence images and the predicted cTNT fluorescence images reached 0.9054 (n=36, p<0.0001) (Figure 9d, e), indicating that our method can accurately assess differentiation efficiency from brightfield images, even for hiPSC-CMs with different morphologies of cardiomyocytes (Figure 9a, b, c).

[0350] In summary, images at the hiPSC-CM stage contain typical features that can significantly indicate differentiation efficiency, and these features can be automatically learned from the data using our proposed method to accurately assess differentiation efficiency from bright-field images.

[0351] 2.2 Deep Learning Method Based on HiPSC-CM Brightfield Images - Pix2pix Model

[0352] We used deep learning to predict cTnT fluorescence signatures from live cell brightfield images and subsequently identify CMs. The pix2pix model, based on a convolutional neural network (CNN), was used for the brightfield-to-fluorescence image conversion task. By training end-to-end on paired brightfield and ground-truth fluorescence images, the model captures the multi-scale characteristics of CMs, enabling it to generate fluorescence predictions for novel brightfield images (Figure 10).

[0353] Table 4 Summary of image data used by pix2pix

[0354] 71 wells (from batches CD03-1 to CD03-6) were randomly divided into training sets (n=35) and test sets (n=36). 62 wells from three additional cell lines (from batches CD03-7 to CD03-9) were used to test the generalization ability of the trained model to new cell lines.

[0355] We prepared a dataset of paired brightfield images and ground-truth (i.e., experimentally obtained) fluorescence images for each well, encompassing a wide range of differentiation efficiencies and different cell lines to increase diversity (Table 4). On the test set, the predicted cTnT fluorescence intensity matched the ground-truth fluorescence intensity at the pixel level, demonstrating that our model can accurately identify CMs (Figure 11a, b, c). Regarding the full-well differentiation efficiency, the Pearson correlation coefficient between the predicted and ground-truth differentiation efficiency indices reached r = 0.93 (P < 0.0001, Figure 11d, e). Furthermore, the trained model could also identify mesoderm in a new dataset of three additional cell lines, with a Pearson correlation coefficient of r = 0.81 (P < 0.0001, Figure 12a, b). Overall, we achieved non-invasive identification of functional cells derived from iPSCs and assessment of differentiation efficiency from brightfield images.

[0356] Example 3 Prediction of differentiation efficiency based on bright field images of hiPSC-CPC stage

[0357] 3.1 The bright field images of hiPSC-CPC cells that have successfully differentiated into the final stage have typical characteristics

[0358] Next, we repeatedly observed the continuously captured image streams and found that in the brightfield images on day 6, the hiPSC-CPC brightfield images corresponding to the areas that eventually successfully differentiated into cTNT-positive cardiomyocytes had a unique texture (Figure 13). Although the texture of these CPCs is very diverse, varying between batches, conditions, and wells, they also have similar morphological characteristics—more three-dimensional and with stronger contrast (Figures 13 and 14). Cells without these characteristics, on the other hand, have a less obvious texture, are relatively flat, and ultimately fail to successfully differentiate, with cTNT staining being negative. These findings were repeatedly verified by the image streams of the cardiomyocyte differentiation process.

[0359] 3.2 Selecting weakly supervised learning to predict differentiation regions based on image labeling methods

[0360] Because hiPSC-CPCs continuously proliferate and migrate during differentiation, we did not use the final cTNT-positive area as a training criterion. Instead, we used the image stream of the entire CM differentiation process to infer CPC regions and generate segmentation masks for the day 6 brightfield images. Annotating mask images requires a certain understanding of CPCs to ensure quality, and due to factors such as the subjectivity of the annotator and the annotation method, the resulting mask images are relatively coarse and cannot achieve pixel-level accuracy. Therefore, we turned to a weakly supervised learning approach, using only classification labels to achieve the goal of localizing hiPSC-CPC regions.

[0361] We segmented full brightfield images into tiles and labeled each tile with ground-truth labels ("0": negative, "1": positive) or uncertain labels ("Uncertain labels") based on the proportion of successfully differentiated regions in the tile (Figures 15 and 16). We first trained the ResNeSt-101 (Zhang et al., 2020) network using a training dataset (n = 8463) consisting of brightfield tiles with ground-truth labels. Next, to localize hiPSC-CPC regions using the trained network, we applied Gradient-weighted Class Activation Mapping (Grad-CAM) (Selvaraju et al., 2017) to generate a coarse localization map to visualize the differentiable hiPSC-CPC regions. To evaluate the performance of our weakly supervised learning framework, we constructed a test dataset of brightfield tiles (n = 12,635, obtained from 35 full brightfield images). Both training and testing included different cell lines (iPS18, iPSB1, H9, etc.), different culture systems (B27 and S12 medium), different starting cell densities, different CHIR treatment doses, and different operators to enhance the network's practical application capabilities (Table 5). The tile-level localization map generated by Grad-CAM and its binarization results were then reconstructed into a complete localization map (Grad-CAM localization map) and a complete binary map (Predicted CPC regions) (Figures 15 and 16).

[0362] Table 5 Summary of image data used in weakly supervised learning

[0363] 126 wells (from batches CD02-1 to CD02-6) were randomly divided into training sets (n=106) and test sets (n=35). The number of negative, positive, and uncertain patches was derived from manually labeled CPC segmentation masks (see Methods). 126 wells from three additional cell lines (from batches CD02-7 to CD02-9) were used to test the ability of the trained model to generalize to new cell lines. CPC segmentation masks for batches CD02-7, CD02-8, and CD02-9 were not manually labeled.

[0364] 3.3 Bright field images in the second stage accurately predict differentiation efficiency

[0365] We analyzed the model's learning performance, showing convergence of the loss curve, with the AUC and ACC curves gradually approaching 1 as the number of epochs increases (Figure 17). The brightfield image patches contain CPC cells with various texture features, and the predictions are accurate. The areas marked as CPC in the binary images predicted from the patches are highly similar to the mask images with manually annotated CPC locations (Figure 18). The areas marked as CPC in the binary images predicted from the full image are highly similar to the mask images with manually annotated CPC locations, with more image details (Figure 19a). We demonstrated the superior performance of the above method through a series of metrics, such as an intersection over union (IoU) of 0.5898±0.1226 and an accuracy of 0.7187±0.1200 (mean±SD, n=33) (Figure 19b). Furthermore, the differentiation regions predicted by the complete binary map were highly consistent with the actual differentiation regions of hiPSC-CMs on day 12. A significant linear correlation was observed between the predicted differentiation efficiency and the actual differentiation efficiency of CMs from the same batch, with a Pearson correlation coefficient (r) as high as 0.88 (n=17, p<0.0001) (Figure 19c, d). Even for the three new cell lines, the Pearson correlation coefficient (r) between the predicted differentiation efficiency and the actual differentiation efficiency of CMs reached 0.83 (n=103, p<0.0001) (Figure 19e, f), demonstrating the model's good generalization ability to new batches.

[0366] The above results show that we can use the bright field images on the sixth day to predict the spatial location of CPC cells that will eventually differentiate successfully and predict the differentiation efficiency in advance.

[0367] We named this series of hiPSC-CPCs with special textures identified by machine learning AI-CPCs.

[0368] 3.4 Image-assisted CPC purification using laser-assisted cell purification

[0369] 3.4.1 Design of Image-Assisted Laser-Assisted CPC Purification Scheme

[0370] Traditional CPC purification methods rely on cell surface markers and require antibody incubation followed by flow cytometry sorting, and cannot be selected based on cell image position information. In order to further purify identified AI-CPCs through machine learning, this experiment used the fluorescent dye DACT-1 (Dual-Activatable Cell Tracker 1) (Halabi et al., 2020) to label cells with fluorescent tags. This dye is in a non-fluorescent state before photoactivation. After entering the cell and irradiated with violet light, a photochemical reaction occurs to form red fluorescent molecules (λmax = 560nm). A confocal microscope with positionable irradiation is used to separate and purify the target cells with the assistance of image information (Figure 20a).

[0371] The specific experimental method is to use DACT-1 to incubate in the dark, and then use confocal microscopy to identify AI-CPC or non-AI-CPC areas. Under the confocal microscope, the designated area (ROI) is restricted to violet light irradiation (λmax = 405nm) to activate DACT-1 molecules in cells in this area. After irradiation, the cells in the irradiated area are visible under 560nm laser and are fluorescently labeled (red fluorescence, RPF) (Figure 20b). Subsequently, the cells are digested and separated into single cells, and then fluorescence-activated cell sorting (FACS) is used to set gates based on the RFP-positive and RFP-negative cell groups, and the two types of cells are collected separately (Figure 20a).

[0372] Finally, count the two types of cells and re-plated them into the culture dish. After they adhere to the wall and continue to be cultured for 3 days, the purification effect can be judged.

[0373] 3.4.2 Excellent purification of CPC and CM, and normal cell status

[0374] Considering the potential for cell damage when irradiating cells with a 405nm laser, the optimal strategy for AI-CPC purification is to select and irradiate the non-AI-CPC region as a ROI. The collected RFP-negative cells are designated as AI-CPCs. Purified AI-CPCs and non-AI-CPCs were cultured in RPMI+B27 medium for 3 days and identified by immunofluorescence using cTNT. The cTNT-positive rate of cardiomyocytes differentiated from purified AI-CPCs was 94.70±3.70% (mean ± SD, n=5), compared with 6.60±4.22% (mean ± SD, n=5) of cardiomyocytes differentiated from the same batch of non-AI-CPCs. The cTNT-positive rate of cardiomyocytes differentiated from the control group without purification was 63.00±11.16% (mean ± SD, n=5) (Figure 21a, b). In contrast, when the AI-CPC region was irradiated with a 405nm laser, the purification effect was acceptable, but due to phototoxicity, the resulting cardiomyocytes were in poor condition, with almost no beating cells visible (Figure 21c, d). Compared with previously reported CPC purification methods, this method significantly improved the purification efficiency. Using the same method, we were able to obtain purified cardiomyocytes based on hiPSC-CM images (Figure 21e, f).

[0375] In summary, we combined artificial intelligence (AI) and laser technology to develop a method to separate cells based on the spatial information of bright-field images, and purify the obtained CPCs or CMs for further downstream applications.

[0376] In addition, the photoactivatable small molecule DACT-1 can be replaced by other toxic photoactivatable probes, and laser irradiation can kill designated cells, eliminating the cell digestion and flow sorting steps, thereby achieving in situ cell purification.

[0377] 3.5 Identification of AI-CPC expressing CPC-related genes

[0378] 3.5.1 Immunofluorescence identification of AI-CPC

[0379] To characterize the biological characteristics of this population of image-recognizing AI-CPCs, we conducted an in-depth analysis of this population of CPCs to determine their specificity and maturity.

[0380] Immunofluorescence results showed that AI-CPCs expressed some known CPCs-specific proteins such as NKX2.5, GATA4, MEF2C and ILS1 on day 6 of differentiation. Under the same conditions, non-AI-CPCs cells outside the AI-CPCs area also expressed related proteins, but the expression level was slightly weaker. Moreover, under conditions of high final differentiation efficiency, a small number of CPC cells from the same batch and treated under the same conditions expressed weakly the classic cardiomyocyte marker protein cTNT on day 6 (Figure 22a, b). Immunofluorescence results of cells that deviated far from normal differentiation conditions (△CHIR≥4) on day 6 showed that NKX2.5, GATA4, MEF2C, ILS1 and cTNT were not expressed.

[0381] These findings suggest that AI-CPCs are a group of correctly differentiated cardiac progenitor cells, with cells that ultimately differentiate into myocardium more efficiently also becoming more mature in the second stage, closer to late-stage cardiac progenitor cells. Currently known CPC marker genes cannot specifically distinguish them.

[0382] 3.5.2 RNA-seq Identification of AI-CPC

[0383] We further characterized AI-CPCs by RNA-seq. The collected samples included AI-CPCs (purified using the DACT-1 method to ensure that cells from the same batch and conditions could ultimately differentiate into beating cardiomyocytes), non-CPCs (ensuring that cells from the same batch and conditions had a final differentiation efficiency of zero), hiPSC-CMs, and hiPSCs, with three biological replicates for each sample.

[0384] PCA analysis and genome-wide heatmap clustering of RNA sequencing (RNA-seq) results showed small intra-group differences and large inter-group differences, indicating good parallelism between the three biological replicates of the same sample and differences in gene expression profiles between different samples (Figures 23a, b). AI-CPCs exhibited gene expression profiles similar to those of classical CPCs, with significant upregulation of NKX2-5, GATA4, MEF2C, TBX5, TBX20, ISL1, HAND1, and HAND2 (Figure 23c). The results also showed that in AI-CPCs, genes associated with the first heart field (FHF) (such as HTBX5, NKX2-5, and HCN4) and the second heart field (SHF) (such as ISL1, NKX2-5, and FLK1) were significantly upregulated, without exhibiting characteristics specific to any particular heart field. Compared with hiPSCs, CM marker genes such as TNNT2, TNNC1, MYH6, and MYH7 were also slightly upregulated in AI-CPCs, but their expression levels were still significantly lower than those in the hiPSC-CM group, consistent with the gene functions enriched by GO analysis (Figure 23d, e).

[0385] It is worth noting that CD82, a cell surface marker previously reported (Takeda et al., 2018), which can be used to sort and purify a group of CPCs whose fate has been determined to differentiate into cardiomyocytes (CM-fated CPCs, CFPs), was not significantly upregulated in our purified AI-CPCs and was even expressed at a lower level than in the non-CPC group (Figure 23c).

[0386] Furthermore, non-CPCs showed upregulation of epicardial cell-characteristic genes, such as WT1 and TBX18, as well as upregulation of fibroblast-characteristic genes, such as COL1A1, COL1A2, VIM, and BMP1 (Figure 23c). This is consistent with previous reports that cardiac fibroblasts are differentiated from epicardial cells (Bao et al., 2017).

[0387] These results indicate that AI-CPCs identified from bright-field images on day 6 possess key molecular signatures of CPCs, but no single gene could independently define them. Cells that failed to differentiate at this stage were more likely to differentiate into cardiac fibroblasts.

[0388] Example 4: Reducing the central area of ​​hiPSC macroclones at the stem cell stage to improve differentiation system efficiency

[0389] 4.1 Discovery of differentiation rules at the edge and center of stem cell clones

[0390] The CD7-captured image stream of the entire differentiation process allowed us to trace back from the final cTNT-positive cardiomyocyte immunofluorescence results to observe the reverse process from cardiomyocytes, cardiac progenitors, cardiac mesoderm, and hiPSCs, allowing us to visually track the location of successfully differentiated cells. During this process, we noticed that hiPSCs located at the edge of the colony at day 0 were more likely to successfully differentiate into hiPSC-CMs, while cells located in the center of the large colony often failed to differentiate (Figure 24a). As can be seen, the cTNT-positive area largely overlapped with the gaps between the 24-hour cell colonies (Figure 24b). We also quantified the specific overlap area and found that 35.7% ± 3.2% (mean ± SD, n = 5) of cTNT-positive hiPSC-CMs were located in areas not covered by cells in the 24-hour brightfield images, a significantly higher proportion than in the control group (18.3% ± 3.6%, mean ± SD, n = 6) (Figure 24c). Based on previously reported findings, this phenomenon may be related to the compactness of cells within hiPSC clones, the sensitivity of the hiPSC clone periphery to WNT signaling (Fred et al., 2016)(Rosowski et al., 2015), and varying cell cycle proportions at different hiPSC confluences (Laco et al., 2018). All of these factors may contribute to the varying responses of hiPSCs to the same CHIR signal. Because these factors are difficult to control, they may also contribute to the instability of cardiomyocyte differentiation batches.

[0391] 4.2 Using Machine Learning to Control iPSC Initial Differentiation State

[0392] The spatially varying differentiation trends within iPSC clones led us to hypothesize that clone morphology might contribute to the differentiation process. Therefore, we established a model to investigate what iPSC starting clone shape leads to optimal differentiation efficiency (Figure 25a).

[0393] To this end, we chose different times after passaging to start differentiation. We introduced iPSC clones of different cell lines and various shapes (Table 6). We quantified their morphological characteristics at 0h (before CHIR treatment) using 343 features of bright field images. For each batch, the final cTnT fluorescence image was collected, and only the wells under the optimal CHIR conditions were considered. The random forest model showed that the standard deviation, minimum value and minimum / maximum ratio of the center point-contour distance, as well as the clone area, perimeter, roundness and convexity were the features most correlated with efficient cell differentiation (Figure 25b,c). The relationship between each individual feature and the final efficiency further showed that initial clones with a moderate area and long and irregular edges tended to have a higher differentiation efficiency (Figure 25d), which is consistent with our observations. Using this random forest regression model, we found that the iPSC differentiation efficiency under optimal CHIR conditions could be predicted based on iPSC morphological characteristics at 0 h, with a Pearson correlation coefficient of 0.76 between the predicted and true values ​​(P < 0.0001) (Figure 25e). This enabled us to monitor iPSC clones in real time by ML to identify the most favorable starting point for differentiation.

[0394] Table 6 Dataset settings used for machine learning-based iPSC clone control.

[0395] 4.3 Adjusting the size of the starting cell clone to improve differentiation efficiency

[0396] The image tracking above reminds us that differentiation efficiency can be improved by adjusting the initial hiPSC clone size. Therefore, during hiPSC passage preparation for differentiation, while maintaining the total cell number constant, by extending the enzymatic digestion time or using repeated pipetting, the clone size was effectively reduced to equivalently increase the length of the clone edge (Figure 26a). The efficiency of cardiomyocyte differentiation from small hiPSC clones reached 91.7% ± 2.9% (mean ± standard deviation, n = 3), the efficiency of cardiomyocyte differentiation from large hiPSC clones was 18.3% ± 7.6% (mean ± standard deviation, n = 3), and the efficiency of cardiomyocyte differentiation from medium hiPSC clones was 48.3% ± 10.4% (mean ± standard deviation, n = 3), which was significantly lower than the differentiation efficiency of small clones (Figure 26b). In summary, we successfully optimized the cardiomyocyte differentiation system by adjusting the initial hiPSC clone size based on the findings from the full cardiomyocyte differentiation image flow analysis. We also found that clone size may be one of the factors that lead to inconsistent differentiation results between batches.

[0397] Example 5: Timely Correction of CHIR Dosage for Bright Field Image Classification in the First Stage of Differentiation

[0398] 5.1 Phase I CHIR Dosage Pattern and Concentration Switching Feasibility Verification

[0399] In this study, we first focused on the local features of differentiated cell images to predict differentiation efficiency and optimize clone size under appropriate conditions. Furthermore, we considered whole-well image features to intervene in early experiments that deviated from differentiation conditions, thereby stabilizing the differentiation system.

[0400] During system establishment and experimental development, we controlled various variables and tested various conditions during cardiomyocyte differentiation. These included the iPSC cell line (iPS18, iPSB1, iPSF, iPSM, H9), starting cell density, iPSC culture medium (mTesR or E8), CHIR concentration and duration, IWR1 concentration (2μM-20μM) and duration, and incubation time at each stage, ultimately identifying key factors influencing differentiation. We clearly found that the CHIR dose in the first stage (from hiPSC to cardiac mesoderm) plays a crucial role in differentiation success, and there is a negative correlation between CHIR concentration and duration within the same batch (Figure 27). Specifically, given the appropriate starting cell density, a difference of just 1μM in the WNT pathway activator CHIR used in the first stage of differentiation can result in a 24-hour difference in the optimal culture medium exchange time. Conversely, if the culture medium exchange time is fixed and a CHIR concentration gradient is designed, a narrow concentration range of 2-4μM CHIR often achieves high differentiation efficiency. This also makes the entire differentiation system highly unstable, particularly when laboratory operators are inexperienced or when using different cell lines. This problem also makes large-scale cardiomyocyte production challenging. This instability may be related to some of the previously mentioned difficult-to-control experimental factors, such as the varying cell cycle proportions of different hiPSC batches and the varying albumin quality across batches. Therefore, we aim to classify the first-stage images to determine whether CHIR is high, medium, or low, allowing for timely adjustments to the CHIR dosage to rescue cells that have taken the wrong differentiation path.

[0401] To achieve the above goals, it is first necessary to verify that the CHIR concentration is switched within 24 hours in the first stage to explore whether it still conforms to the dose-effect law. As shown in the figure, the dose-effect law is still followed after switching the concentration (Figure 28a, b). For example, if CHIR 4μM is used for 0-48, the dosage is obviously too low and the differentiation efficiency is not high. However, if the CHIR concentration is adjusted to 6 or 8μM within 24 hours, the differentiation efficiency is significantly improved (Figure 28a). The above results verify the feasibility of using early images to judge the high, medium and low levels of CHIR and to adjust the CHIR concentration in a timely and early manner.

[0402] 5.2 Design ideas and feature selection for the first stage image CHIR high and low classification

[0403] The first-stage cardiomyocyte image classification system we proposed consists of a feature extraction module and a machine learning classification module: A live cell bright-field image stream of a well from 0 to 12 hours is input. The feature extraction module first calculates its high-dimensional feature representation, and then the machine learning classification module infers the category to which its concentration belongs ("low", "moderate" or "high").

[0404] To train and validate this classification system, we prepared a dataset consisting of streams of whole-well bright-field images (n = 384) (Table 7), which encompassed different influencing factors (cell line, batch, initial cell density, CHIR dose, etc.); the dataset was then randomly divided into a training set (n = 268) and a test set (n = 116). To add concentration category labels to the dataset, for each batch, we determined the optimal CHIR concentration range based on the final differentiation results (cTNT immunofluorescence images) under given CHIR durations (24 h, 36 h, and 48 h), and calculated the "ΔCHIR concentration" for other concentration levels to measure their deviation from the optimal range; thus, all wells in the dataset were labeled according to their CHIR concentration: low (ΔCHIR concentration < 0), moderate (ΔCHIR concentration = 0), and high (ΔCHIR concentration 0) (Figure 29a, b).

[0405] To enable the classification system to distinguish between different well types, we needed to select features from the brightfield image stream from 0 to 12 hours in the first phase. Analysis of the time-series brightfield images from the first phase revealed the following overall observations: After CHIR addition at 0 hours, the area of ​​the hiPSC colonies decreased continuously, with the rate of reduction possibly related to the CHIR concentration and the size of the hiPSC colonies. The image contrast at the colony edges increased, the colony color gradually deepened, and the internal texture changed. Dead cells gradually became visible in the high CHIR group.

[0406] Based on the above observations, we designed a feature set consisting of 21 variables to perform the classification task, including fractal dimension, cell coverage statistics (area, perimeter, area-perimeter ratio, brightness, local entropy), and optical flow. (Texture features were also explored but appear irrelevant to classification; data are not shown here.) Of these features, "optical flow" was calculated for every two consecutive timestamps (this type of feature is designated as Type-II), while the others were calculated for each timestamp (this type of feature is designated as Type-I) (Figure 29c); in both cases, a real number sequence was generated to represent the feature value. Furthermore, this experiment also normalized the values ​​of "area," "perimeter," "area-perimeter ratio" ("AC ratio"), and "optical flow" by dividing them by the first value in the sequence (referred to as "relative features"); the other features were used unnormalized (referred to as "absolute features"). Finally, timestamps T1-T10 were divided into early, middle, and late stages, and the features of each stage were averaged (Figure 29c). Therefore, each of these seven features gives three real numbers (corresponding to prophase, metaphase, and anaphase), resulting in a 21-dimensional feature representation for each well. They can reflect the state of the cell and its response to different CHIR concentrations; in this way, the image flow can be described by a 21-dimensional vector representation.

[0407] Table 7 Summary of the first-stage CHIR dose classification dataset. 384 wells (from batches CD01-1 to CD01-4) were randomly divided into training sets (n=268) and test sets (n=116). For each batch, the CHIR concentration with an average cTnT+ cell percentage ≥20% was labeled as optimal, while concentrations outside the optimal range were labeled as low or high.

[0408] 5.3 Based on 0-12h brightfield image machine learning, CHIR can be classified into three categories: high, medium and low

[0409] To visualize this 21-dimensional feature space, we projected it onto the most discriminative two-dimensional plane using linear discriminant analysis (LDA) (Hastie et al. 2009) (Figure 29b). We found that the three concentration classes were clearly separated, indicating that the 21 variables we extracted indeed contained the necessary information for subsequent classification (Figure 30a). We then trained a logistic regression classifier on the training dataset to automatically predict the CHIR concentration class from the extracted image features. It achieved high accuracy on the test dataset (93.1%, 84.5%, and 78.4% for CHIR durations of 24, 36, and 48 hours, respectively) (Figure 30b). This suggests that our classification system captures the underlying relationship between the first-stage brightfield images (0-12 hours only) and final differentiation efficiency purely from the data. However, visualization using PCA (Figure 30c) reveals that using the entire 21-variable feature set would contain a significant amount of information irrelevant to classification. Therefore, we next considered reducing the dimensionality of the feature representation through variable selection to make the classification system more robust. We performed a one-way analysis of variance (ANOVA) on all 21 variables in the training set (n = 268, with the classification labels subscripted for a 24-hour CHIR duration) and ranked the variables based on their p-values ​​(Figure 30f). The four variables with the lowest p-values—late-stage optical flow, cell brightness, and colony perimeter—and mid-stage cell brightness were selected as the final feature set. The brightfield image streams of each well were then mapped into a 4-dimensional feature representation. These four variables may be interpreted as follows: optical flow measures the speed of cell movement, cell brightness is related to the compactness of hiPSC colonies, and colony perimeter reflects the size and density of the colonies, all of which may influence the subsequent differentiation direction of the cells. We again visualized the 4-dimensional feature space using PDA (Figure 30d) and LDA (Figure 30e). We found that using only the 4-dimensional feature vector still largely retained the ability to distinguish between different concentration classes; pores within the same class became more concentrated, while pores between classes became more separated. We repeated the variable selection process for the annotations of CHIR durations of 36 and 48 hours. Even using only the four selected variables, the classifier still achieved a fairly high accuracy (Figure 30g).

[0410] 5.4 Cross-validation of different batches predicts that the CHIR bias is basically correct

[0411] To test the transferability of the classification system across different differentiation batches in real-world applications, we performed cross-batch cross-validation to simulate this scenario. In the cross-validation experiment, we annotated each well under the condition of a 24-hour CHIR duration and trained the classifier on three batches and tested it on new batches. When testing on new batches, predictions for image streams with the same CHIR concentration were aggregated into a single "bias score" ranging from -1 (very likely "low") to +1 (very likely "high"). The classifier's predictions were highly consistent with the true labels; in particular, when testing on CD01-1 and CD01-3, the classifier's estimated bias scores increased from negative to positive values ​​as CHIR concentration increased (Figure 31).

[0412] Cross-validation results demonstrate the potential of our method to learn general, batch-independent classification criteria from data, enabling prediction of actual CHIR concentrations across unseen batches, thereby stabilizing cardiomyocyte differentiation efficiency. Future studies using large numbers of high-quality brightfield images from different batches will help us discover new, more adaptable and robust features.

[0413] Example 6 Image-Assisted Small Molecule Screening to Optimize Cardiac Differentiation System

[0414] 6.1 The high-dose CHIR group tended to differentiate into somite mesoderm

[0415] We further optimized the cardiomyocyte differentiation system through a first-stage small molecule screening. To understand the effects of high or low CHIR doses on cell fate determination in the first stage, we performed RNA-seq sequencing on cells in the first stage at different CHIR concentration and treatment time combinations. Samples were collected during the first stage of differentiation (0-72 hours). A total of 10 cell samples were collected at different CHIR doses (hiPSCs; CHIR 2μM 48 hours, 6μM 24 hours, 6μM 36 hours, 10μM 24 hours, 8μM 36 hours, 6μM 48 hours, 12μM 24 hours, 12μM 36 hours, and 10μM 48 hours), including three groups with low, moderate, and high CHIR doses. Each group had three replicate wells with the same conditions to determine its differentiation efficiency.

[0416] PCA analysis and genome-wide heatmap clustering of RNA sequencing (RNA-seq) results showed that among the nine samples with different CHIR doses, samples with successful differentiation clustered closely together, with the high or low CHIR dose groups surrounding the moderate dose group (Figure 32a, b). Stemness genes in hiPSC samples were normally expressed, but with increasing CHIR concentration or duration, stemness genes, including NANOG, POU5F1, OTX2, and HESX1, were gradually downregulated. In the moderate CHIR dose group (i.e., the successful differentiation group), cardiac mesoderm-related genes, including MESP1, MESP2, and EOMES, were significantly upregulated. In the high CHIR dose group, somite mesoderm-related genes, including CDX1, CDX2, MSX1, and MSGN1, were significantly upregulated (Loh et al., 2016) (Figure 32c, d).

[0417] 6.2 Knockdown of somite mesoderm genes at high CHIR doses still results in myocardial differentiation

[0418] Excessive addition of CHIR during the first differentiation stage blocks cardiomyocyte fate, redirecting differentiation toward somite mesoderm. Therefore, we attempted to knock down somite mesoderm genes, including CDX1, CDX2, MSX1, and MSGN1, in hiPSCs. Knockdown of CDX2 and MSX1 allowed cells to continue to differentiate toward cardiomyocytes despite high-dose CHIR treatment during the first differentiation stage (Figures 33a, b, c; CDX2 knockdown results not shown). Knockdown of CDX1 and MSGN1 had no significant effect on the range of CHIR application in cells (results not shown).

[0419] 6.3 CPC stage image-assisted CHIR high-dose small molecule screening

[0420] On this basis, we are committed to using small molecules to achieve the above effects, so that hiPSCs can still maintain the correct differentiation direction in the high CHIR dose group, thereby expanding the applicable range of CHIR concentration and time, and improving the efficiency and stability of the myocardial differentiation system. Using the above-mentioned AI-CPCs image learning method using weakly supervised learning, we have used hiPSC-CPC brightfield images to more accurately predict the efficiency of final differentiated cTNT-positive cardiomyocytes. Therefore, for the small molecule screening results, we only collected brightfield images on the 6th day of differentiation under different small molecule treatments, input them into the previously trained weakly supervised learning network, and combined with Grad-CAM to predict differentiation efficiency. Compared with the traditional use of cTNT immunofluorescence or cell lines with established cTNT reporter systems as screening criteria, this method significantly shortens the screening cycle and saves manpower and material resources.

[0421] The small molecule screening work used a small molecule library including more than 3,000 compounds, and the differentiation experiment was carried out in a 384-well plate. Differentiation began when the hiPSC density was appropriate. Under the condition of high CHIR concentration, the small molecules to be screened were added from 0 to 48 hours (the initial concentration was uniformly 2 μM), and CHIR and the screening small molecules were removed at the same time at 48 hours. The subsequent differentiation process was normal, and the bright field images of each well on the 6th day were collected. Due to the instability of myocardial differentiation, each batch was set up with a pair of wells to ensure that the group without screening small molecules and the group with high CHIR dose could not differentiate into myocardium normally (negative control, NC), and the group with normal CHIR dose had normal differentiation (positive control, PC). According to the previous image weak supervision learning method, the bright field images on the 6th day were preprocessed, predicted, and the differentiation efficiency was predicted (Figure 34a). The first round of screening of effective small molecules (Hitcompounds) was followed by effect verification, concentration adjustment, testing on different cell lines, and testing of small molecules with the same target (Figure 34b).

[0422] Through a weakly supervised learning model based on brightfield images, we effectively screened compounds that can maintain correct myocardial differentiation at high CHIR concentrations, successfully expanding the application range of CHIR concentrations and further stabilizing and optimizing the hiPSC-to-myocardial differentiation system.

[0423] This study first established a commonly used hiPSC-to-cardiomyocyte differentiation system, progressing through the hiPSC, mesoderm, cardiac progenitor, and cardiomyocyte stages. Continuous live-cell brightfield image streams were captured throughout the differentiation process of multiple batches of different stem cell lines, along with final cTNT immunofluorescence results to evaluate differentiation efficiency. Through machine learning of images from the entire differentiation process, solutions were proposed to address instability issues in hiPSC-CM differentiation at each stage from the following perspectives, while also optimizing the differentiation system (Figure 35):

[0424] 1) Quality control of starting cells during the hiPSC stage: Reverse tracking of bright-field images throughout the cardiomyocyte differentiation process revealed that ultimately differentiated cardiomyocytes were more likely to be located at the edges of the hiPSC clones, while differentiation often failed in the center of the clones. Further experiments confirmed that smaller starting hiPSC clones facilitated efficient cardiomyocyte differentiation. This may also be a contributing factor to system instability.

[0425] 2) Intervention of differentiation direction in the early stage (differentiation into mesoderm) from 0 to 72 hours: The concentration and treatment time of CHIR are crucial to differentiation efficiency and are unstable across batches. This study found that the concentration of CHIR is negatively correlated with the treatment time, and verified that early switching of the CHIR concentration can still lead to successful differentiation. Therefore, machine learning was performed on the features of the bright field image stream in the first stage of differentiation, and the actual CHIR concentration used in the batch (low, moderate, or high) was successfully determined 12 hours after differentiation. By scoring the differentiation conditions in the early stages of differentiation, timely intervention and rescue of incorrectly differentiated cells can be achieved, returning to the correct myocardial differentiation path.

[0426] 3) Predicting differentiation efficiency in the middle and late stages of differentiation (hiPSC-CPC and hiPSC-CM stages): For the final hiPSC-CM stage, this study developed a deep learning method combining GoogLeNet with CycleGAN to predict differentiation efficiency from brightfield images to cTNT fluorescence images. This study also used a weakly supervised learning method to perform image learning on CPC regions that successfully differentiated on day 6 and exhibited unique image features. This group of AI-CPCs was successfully identified and differentiation efficiency was predicted in advance. Since cells that differentiated incorrectly at this stage cannot be corrected, timely control can be achieved based on the predicted differentiation efficiency.

[0427] 4) Purification of differentiation intermediates: Based on the above-mentioned image recognition, combined with the light-activated small molecule DACT-1 and microscope laser technology, AI-CPC and other misdifferentiated cells can be purified to further improve differentiation efficiency.

[0428] 5) Image-based small molecule screening to stabilize the system: In the first phase, small molecules were screened at high CHIR doses from 0 to 48 hours. Bright-field images of live cells were captured on day 6 of differentiation and fed into a weakly supervised learning network to predict differentiation efficiency. Ultimately, the addition of Compound A enabled cells to differentiate efficiently even in the high CHIR dose group, significantly broadening the applicable CHIR concentration range, optimizing the differentiation system, and enhancing stability.

[0429] This article is the first to combine label-free bright-field dynamic imaging of cells with machine learning to stabilize and optimize the myocardial differentiation system from multiple perspectives, providing methods and new ideas for the efficient, stable, and large-scale production of cardiomyocytes differentiated from induced pluripotent stem cells, and providing guarantees for in vitro response or cell therapy of cardiomyocytes.

[0430] Example 7: Transferring Machine Learning Strategies to Kidney and Liver Differentiation

[0431] The success of machine learning in regulating and optimizing cardiomyocyte differentiation encourages us to transfer this strategy to other iPSC differentiation processes, such as kidney and liver cells, which may also be valuable for cell-based therapies or drug toxicity assessment.

[0432] 7.1 Concentration Assessment in Early Stages of Renal Differentiation

[0433] During the early differentiation of iPSCs into kidney organoids, the optimal CHIR concentration is crucial for high differentiation efficiency but fluctuates between batches, depending on the cell line, passage number, and culture conditions (Figure 36a). However, cells treated with different CHIR concentrations (low, optimal, and high) exhibit distinct bright field image characteristics (loose, normal, and dense, respectively) on day 4 (when CHIR is removed) (Figure 36b). Next, we investigated whether CHIR concentration could be assessed by ML on day 4.

[0434] We prepared a dataset of day 4 brightfield images of different cell lines (iPS-B1, iPS-F, iPS-M, H9, WIBR3) and CHIR concentrations ranging from 3 to 16 μM. To evaluate CHIR concentration, day 4 brightfield images were labeled as low, optimal, or high based on their day 4 signatures and day 9 immunofluorescence staining for SIX2 (a marker for renal progenitor cells, NPCs) (Figure 36c). t-SNE plots of SIFT local features extracted from the brightfield images demonstrated clear separation between the different CHIR concentration groups (Figure 36d). Using these local features, a trained logistic regression model accurately classified the brightfield images in the test set with 98.97% accuracy (Figures 36e,f). Because cells with dense morphology should be terminated early, while cells with loose morphology can achieve efficient differentiation with prolonged CHIR treatment, early assessment of CHIR concentration provides valuable guidance for stabilizing the renal differentiation system.

[0435] 7.2 Identification of Hepatic Definitive Endoderm Cell Regions

[0436] Low reproducibility of differentiation efficiency across different batches is also a key challenge in liver differentiation systems. Therefore, we explored the application of machine learning to non-invasively identify definitive endoderm (DE) cell regions in brightfield images (72 hours, the first stage of liver differentiation) for early assessment of liver differentiation status and potential subsequent image-based cell purification (Figure 37a). We captured live cell brightfield images at 72 hours and corresponding immunofluorescence images of SOX17 (a DE marker gene); we introduced differential differentiation efficiency by adjusting the dosage of small molecules (CHIR and IDE1) used in different cell lines during the first stage of differentiation (Figure 37b). We then trained a weakly supervised learning model on the DE brightfield images, requiring only the full-image class label (i.e., "positive" or "negative," based on the proportion of SOX17+ cell regions; see Methods). After training, the trained model predicted endoderm cell regions that closely matched the SOX17 fluorescent label (Figure 37c). The predicted proportion of endoderm cell areas was also correlated with the proportion of true SOX17+ cell areas (Pearson's r=0.92, P<0.0001) (Figure 37d). These two extended applications further validated the versatility of our strategy.

Claims

1. A neural network model for predicting and / or determining the efficiency of differentiation of starting cells into target cells, which is obtained by the following steps: A bright field image of cells at a specific stage of differentiation is provided as an input image, and a corresponding target cell image confirmed by target cell-specific staining is used as a correct image, and a neural network is used for learning to obtain the neural network model.

2. The neural network model of claim 1, wherein the neural network comprises (1) an image classification neural network, and (2) an image conversion neural network.

3. The neural network model of claim 1 or 2, wherein the starting cells are pluripotent stem cells, such as embryonic stem cells (e.g., embryonic stem cells no older than 14 days) or induced pluripotent stem cells.

4. The neural network model of any one of claims 1 to 3, wherein the target cells are differentiated cells, for example, the cells are selected from neurons, skeletal muscle cells, hepatocytes, kidney cells, fibroblasts, osteoblasts, chondrocytes, adipocytes, endothelial cells, interstitial cells, smooth muscle cells, cardiomyocytes, nerve cells, hematopoietic cells, and pancreatic islet cells.

5. The neural network model of any one of claims 2 to 4, wherein the (1) image classification neural network is selected from googleNet, VGG, ResNet, ResNeXt and SE-Net, preferably googleNet.

6. The neural network model of any one of claims 2 to 5, wherein the (2) image conversion neural network is selected from CycleGAN, DiscoGAN and DualGAN, preferably CycleGAN.

7. The neural network model of any one of claims 2-6, wherein the (1) image classification neural network is googleNet, and the (2) image conversion neural network includes two CycleGANs.

8. The neural network model of claim 7, wherein GoogleNet classifies the brightfield image tiles into class "0" and class "1", and then inputs the brightfield image tiles and the corresponding stained image tiles into CycleGAN-0 and CycleGAN-1 for learning, respectively.

9. The neural network model of claim 1, wherein the neural network comprises a pix2pix model.

10. The neural network model of claim 9, wherein the pix2pix model comprises a generator G that learns to predict stained images from brightfield images, and a discriminator D that learns to distinguish between true and false brightfield-fluorescence image pairs.

11. The neural network model of claim 1 , wherein the neural network is a random forest regression model.

12. The neural network model of any one of claims 1 to 11, wherein the following features of the bright field image are used to quantify the morphological characteristics of the cells: (17) local entropy, cell brightness, cell contrast, and total variation; (18) Hu invariant moments 1-7; (19) SIFT 1~256; (20)ORB 1~64; (21) Area, perimeter, and area / perimeter ratio; (22) Solidity, convexity, and roundness; (23) Maximum center-to-contour distance (CCD), minimum CCD, minimum / maximum CCD ratio, mean CCD, standard deviation of CCD; and / or (24) Spacing.

13. The neural network model according to any one of claims 1 to 11, wherein the specific stage of differentiation is the final stage of induced differentiation.

14. The neural network model according to any one of claims 1 to 11, wherein the specific stage of differentiation is an intermediate stage of induced differentiation.

15. The neural network model according to any one of claims 1 to 11, wherein the specific stage of differentiation is an initial stage of induced differentiation.

16. The neural network model of any one of claims 1 to 15, wherein the target cell-specific staining is immunofluorescence staining.

17. A neural network model for predicting and / or determining regions of cells capable of differentiating into target cells during differentiation of starting cells into target cells, the model being obtained by the following steps: A bright field image of cells at a specific stage of differentiation is provided as an input image, and a corresponding image of cells suspected of being able to differentiate into target cells is used as a correct image. A neural network is used to perform weakly supervised learning to obtain the neural network model, wherein the neural network includes (1) an image classification neural network, and (2) an image positioning neural network.

18. The neural network model of claim 17, wherein the starting cells are pluripotent stem cells, such as embryonic stem cells or induced pluripotent stem cells.

19. The neural network model of any one of claims 17-18, wherein the target cells are differentiated cells, for example, the cells are selected from neurons, skeletal muscle cells, hepatocytes, kidney cells, fibroblasts, osteoblasts, chondrocytes, adipocytes, endothelial cells, mesenchymal cells, smooth muscle cells, cardiomyocytes, nerve cells, hematopoietic cells, and pancreatic islet cells.

20. The neural network model of any one of claims 17 to 19, wherein the (1) image classification neural network is selected from Resnet-101, VGG, ResNeXt, SE-Net, preferably Resnet-101.

21. The neural network model of any one of claims 17-20, wherein the (2) image localization neural network is selected from Grad-CAM.

22. A method for predicting and / or determining the efficiency of differentiation of a starting cell into a target cell, the method comprising: (1) Acquire bright-field images of cells at specific stages of differentiation; (2) analyzing the bright field image using the neural network model for predicting the efficiency of differentiation of starting cells into target cells according to any one of claims 1 to 16; (3) Determine the differentiation efficiency.

23. The method of claim 22, wherein the differentiation efficiency is quantified by a differentiation index (or differentiation efficiency index), wherein For an MxN fluorescence staining image I (intensity value ∈ [0, 1]), its “differentiation efficiency index” is defined as the total fluorescence intensity of pixels whose intensity value exceeds the threshold α, that is, Where M and N are the height and width of the fluorescence image.

24. A method for predicting a cell region capable of differentiating into a target cell during differentiation of a starting cell into a target cell, the method comprising: (1) Acquire bright-field images of cells at specific stages of differentiation; (2) analyzing the bright field image using the neural network model for predicting the region of cells capable of differentiating into target cells during differentiation of starting cells into target cells according to any one of claims 17 to 21; (3) Identify the cell regions that can differentiate into target cells.

25. A method for isolating and / or purifying cells at a specific stage in the process of differentiating from starting cells to target cells, the method comprising: (1) Acquire bright-field images of cells at specific stages of differentiation; (2) analyzing the bright field image using the neural network model for predicting the region of cells capable of differentiating into target cells during differentiation of starting cells into target cells according to any one of claims 17 to 21; (3) identifying regions of cells capable of differentiating into target cells; (4) treating cells with a laser-activated probe such as DACT-1; (5) treating cells outside the region of cells determined to be capable of differentiating into target cells with a laser, and (6) Cells within the cell region determined to be capable of differentiating into target cells are sorted out.

26. The method of claim 25, wherein the sorted cells have an increased ratio of differentiation into target cells.

27. The method of claim 25, wherein the laser-activated probe is a toxic laser-activated probe.

28. The method of any one of claims 25-27, wherein the target cell is a cardiomyocyte and the stage-specific cell is a cardiac progenitor cell.

29. A method for screening conditions that can promote differentiation of starting cells into target cells, the method comprising: 1) Changing one or more differentiation conditions at a specific stage of differentiation; 2) predicting / determining the differentiation efficiency under the changed differentiation conditions by the method of claims 22-24; 3) Determine the conditions under which the differentiation efficiency is optimal as the conditions that promote differentiation.

30. The method of claim 29, wherein the differentiation condition is contact with a given small molecule compound to be tested, for example, differentiation is performed in a culture medium containing the given small molecule compound to be tested.

31. The method of claim 29, wherein the target cell is a cardiomyocyte.

32. The method of claim 31, wherein the specific stage of differentiation is a stage in which pluripotent stem cells differentiate into cardiac mesoderm.

33. The method of claim 31 or 32, wherein the differentiation condition is adding a small molecule compound to be tested at a given CHIR99021 concentration.

34. A method for differentiating cardiomyocytes from pluripotent stem cells, such as embryonic stem cells (e.g., embryonic stem cells no older than 14 days) or induced pluripotent stem cells, the method comprising: 1) At the pluripotent stem cell stage (differentiation initiation stage), using the method of any one of claims 22 to 24 to predict and / or determine the differentiation efficiency, thereby performing quality control on the starting pluripotent stem cells; 2) At an early stage of differentiation (e.g., mesoderm stage), using the method of any one of claims 22-24 to predict and / or determine differentiation efficiency, thereby evaluating early differentiation conditions and maintaining or modifying the differentiation conditions accordingly; 3) At a mid-to-late differentiation stage (such as a cardiac progenitor cell CPC or cardiomyocyte CM stage), using the method of any one of claims 22-24 to predict and / or determine the differentiation efficiency, thereby terminating the differentiation or continuing the differentiation accordingly; and / or 4) Purifying differentiated intermediate cells capable of differentiating into cardiomyocytes based on the method of any one of claims 25 to 28, thereby improving differentiation efficiency.

Images