Prediction of Stem Cell State Using Cell Image-Based Deep Learning

ESC-and-iPSC_detection.git.

Stem cells represent an ideal source for regenerative medicine; however, longitudinal assessment of stem cell phenotype and function is challenging. Contrastingly, a convolutional neural network (CNN) algorithm can automatically extract the image features and produce highly accurate image recognition. Thus, this study implements CNN to establish stable and reproducible cell culture experiments by predicting a unique morphology of pluripotent stem cell (PSC) lines. Interestingly, the algorithm distinguishes the PSC lines cultured in the different cell culture conditions, such as the presence or absence of small molecules and/or the long-or short-term culture in our induced PSC (iPSC) models, which include iPSC lines with abnormal gene expression patterns and genomic abnormalities. Our deep learning technology accurately classifies the various cell lines with or without genetic defects using only the cell images, without any labeling process. This suggests that the CNN system may simplify the various tasks involving stable cell cultures and their differentiation.
biomedicine and cell biology in the near future. Recent papers have emphasized the potential use of deep learning in cell and molecular biology, with some of the effective applications documented to include protein-protein interactions, cardiology, and cardiac imaging, alongside drug-target interactions. [24][25][26][27] The experimental tests that require visual pattern recognition might soon undergo a significant transformation based on what has been accomplished thus far with deep learning. [28] Such that, CNN technology can now be applied in the field of stem cells. Indeed, while many studies using stem cells have suggested that CNN can detect morphological changes according to the cell culture conditions, there have been actual studies on this. Therefore, it is necessary to determine whether the cell culture conditions using the numerous stem cell lines are applicable to CNN technology. In addition, it is important to determine which culture medium the trained CNN used simply by observing the cell morphology.
As a result, we first discovered the feasibility of CNN technology using mouse ESC lines cultured in various culture medium conditions. To apply another PSC lines to CNN, we used various mouse iPSC lines with genetic stability according to the presence or absence of small molecules. As a consequence of the training, CNN could accurately distinguish the subtle differences in cell morphologies with high accuracy. These findings indicate that CNN technology may be used as a tool to distinguish differences of cell morphology in images as well as predict the levels of genetic safety, which suggests that AI-based precise recognition of cellular morphology under light microscopy may have a considerable influence on how future cell assays are performed.

ResNet-50 Algorithm Structure
Initially, an algorithm was constructed to test whether the different types of cell images could be trained via CNN. For CNN training, mESCs (B6 and J1) and miPSCs were cultured in three different types of cell culture conditions, respectively. The culture medium conditions are as follows: in ESC, leukemia inhibitory factor (LIF)-added medium to maintain the pluripotency of stem cells, LIF-free medium, and insulin/transferrin/selenium (ITS) medium to induce differentiation. In iPSC, LIF and four small molecules (4SMs: SB431542, PD0325901, thiazovivin, and ascorbic acid (AA)) medium that maintains the genetic stability in iPSCs, [29] LIF and 4SMs-free condition and ITS condition. The cells were collected at different times (1, 3, 6, 12, and 24 h after changing each different medium condition) from the transmitted light microscopy and each cell image was used as input data in the CNN algorithm ( Figure 1A). The ResNet-50 algorithm, which consists of 50 deep layers, is most commonly applied to analyzing visual images. [30] The image features of input data are extracted through convolution layers and pooling layers ( Figure 1B). The size of the input data of the CNN model increased from 240 Â 320 to 246 Â 236 temporarily by zero-padding processing. Accordingly, an image size of 120 Â 160 with 64 channels is input to the convolution layer 1, before a 60 Â 80 image size with 64 channels is input to the second neural network by a pooling operation in the pooling layer. In the last fully connected layer, the feature map extracted by the filters is applied to the activation function. While the activation function used is ReLU, which does not cause gradient vanishing problems and has the advantage of fast learning. [31] Afterward, the training was completed by classifying images with a softmax function. [32] To verify that CNNs were properly trained, unlabeled cell images were applied to trained CNNs to confirm separability under which of the two culture medium conditions they were cultured. More detailed information on the ResNet-50 structure can be found in the Supplementary data.

Inaccurate Analysis with an Initial 1000 Training Images in ESCs
The LIF and ITS medium conditions were applied to effectively train the CNN model. Since ITS medium significantly induced cell morphology within the initial 24 h, it has been simple to compare the cell morphological differences between LIF and ITS conditions. Initially, 1000, 800, and 100 images for each medium condition were used as the train, validation, and test set, respectively. We cultured B6 mESCs in LIF and ITS medium condition ( Figure 2A). It is evident that cells proliferated rapidly after 12 h in the ITS condition and after 24 h in LIF condition. CNN training results show that overfitting occurred in which the validation loss value was not constant and fluctuated at 1, 12, and 24 h graphs ( Figure S1, Supporting Information, left graph). The first cause of overfitting is that the number of training images and of training epochs are insufficient. Second is that it is difficult to image in the edge of cells when cell proliferation is too fast. The trained network with 1000 training images was not able to predict with high accuracy in both the training and validation groups. Confusion matrix in the right showed how well-trained CNNs could distinguish them with high accuracy using 100 images per media condition. With the exception of 3 h, accuracy was not very high, hence 1000 training images are not appropriate for CNN training ( Figure S1, Supporting Information, right table). Therefore, we have attempted to reset the ideal number of images to increase accuracy.

Accurate Analysis with 2000 Training Images in ESCs
Next, the number of training images was increased to 2000 to improve the network performance, while the numbers of validation and test images remained constant. As a result of training with 2000 images, the accuracy of training and validation was close to 1, and the loss of them was nearly 0, indicating that the training was well-conducted ( Figure 2B). Comparing CNN training results, training accuracy increases as the number of training images increases ( Figure 2C). Therefore, the ResNet-50 algorithm, which can predict different cell morphologies under each cell culture condition, chose to use 2000 training images rather than 1000 images. Next, the accuracy with which CNN distinguished the difference in cell morphology, in terms of precision and recall, was investigated. Precision is the proportion of what is classified as "true" in model learning, while recall is the proportion of what the model predicts to be true among what those previously classified as "true," [33] In both the LIF(þ) and ITS groups, the precision and recall values used in 2000 images were close to 1 in all time zones ( Figure 2D). Therefore, we confirmed that CNN can distinguish cell types that change under various media conditions with 2000 training images.
If the experimenter accidentally uses different media conditions, it can be challenging to identify which cells are cultured in the incorrect media composition by observing them with only human eyes. We applied LIF-added (LIFþ) and LIF-free (LIFÀ) media conditions to mESCs to induce subtle variations in cell morphology and confirmed whether CNN can distinguish them. In fact, it is evident that there is little difference in cell morphology between the two conditions, making it impossible to determine which media conditions it was cultured under solely by examining the cell shape ( Figure 3A).
Surprisingly, training results unveiled that CNN could detect minute morphological variations in cells under two different cell culture conditions. The accuracy of training and validation is close to 1, and the loss value of them is close to 0 by time ( Figure 3B). Notably, the accuracy at 24 h was slightly reduced because the cell size was too large to accurately fit within one image; however, the accuracy remained higher than 90%. The precision and recall values, which are indicators of CNN performance, were close to 1 with LIF(þ) and LIF(À) conditions ( Figure 3C). These results show that CNN can distinguish subtle differences in cell types that are not visible to the human eye, suggesting that CNN may predict the cell culture conditions used during stem cell culture using only the cell images.
Another mESC line, the J1 mESC was also photographed at a single time point under the same culture conditions, allowing the training method to be performed under the same conditions ( Figure S3A, Supporting Information). CNN completed the training with high accuracy irrespective of media conditions ( Figure S3B,C, Supporting Information). In addition, in the test for evaluating the model, CNN was able to distinguish the conditions of the culture solution by analyzing the cell images that are difficult to discern with the human eyes. There is actually no statistically significant difference between B6 and J1 mESCs, according to a quantitative analysis of how the cell shape varies depending on the two media conditions ( Figure S2A, Supporting Information).

Visualizing the Activation Layers with Convolutional Neural Network (CNN)
Cell images used for CNN training are analyzed to find out "how" features are extracted in each layer and "which" features are used to train in CNN. Deep neural networks (DNNs) require www.advancedsciencenews.com www.advintellsyst.com enormous amounts of computation in a series of layers called hidden layers. The visualization of the activation of each layer corresponding to a CNN consists of characteristic maps that are output by several layers of convolution and pooling in the network given the corresponding input data. The next illustrates the way in which the filters of each layer, learned by the network, analyze the input data ( Figure 4). The B6 mESC images cultured in each cell culture condition are an input layer size of 240 Â 320 size with three channels (depth). CNN analyzes and extracts information from the yellow-activated regions of the cell images in each layer. As the convolutional layer (Conv) becomes deeper from Conv1 to Conv5, the size of the activation layer gradually decreases and the channels (depth) increases. A decrease in the size of the activation layer means a decrease in the number of pixels in the cell image, which changes the cell image abstractly. However, as the channel increases, the number of analysis Figure 3. Results of CNN training with LIF(þ) and LIF(À) conditions. A) Minimal differences in cell morphology are observed with LIF(þ) and LIF(À) conditions. Scale bars, 50 μm. B) Training accuracy is close to 1. In the confusion matrix results, the accuracy of the LIF and the ITS group is over 93% and 90%, respectively. C) The values for precision and recall show close to 1 for 24 h. However, the value is slightly lowered at 24 h due to cell size and debris.
www.advancedsciencenews.com www.advintellsyst.com iterations increases. Irrelevant information is filtered, and useful information is highlighted and enhanced through repeated transformations of the original cell image input into the DNN. The weights of the final activation layer are fed into the softmax function and the iterative relationship of these values for all the images applied to the CNN provides a pattern that helps identify the images.

The Distinction of Shape Differences in IPSCs According to the Presence or Absence of Small Molecules
The iPSCs can be differentiated into all three germ layers, similar to ESCs, and be used to treat various diseases. However, iPSCs exhibit abnormal gene expression and genomic abnormalities, raising concerns about the safety of their use in clinical applications. [9][10][11][12] In our previous study, several small molecules were found to maintain genomic stability during the long-term culture of stem cells or via reprogramming somatic cells to iPSCs in vitro. [29] Four small molecules that we discovered were the MEK inhibitor (PD0325901), TGF-β inhibitor (SB431542), thiazovivin, and AA, all of which can improve the reprogramming efficiency, differentiation potentials, and genetic stability. Subsequently, we treated the miPSC lines (miPSC(þ) line 1, miPSC(þ) line 2) with these 4SMs alongside the miPSC lines (miPSC(À)) that were not treated. Both types of miPSCs were at a middle passage (P15). The ability of the CNN to distinguish any cell morphology changes dependent upon the effect of the LIF and 4SMs was demonstrated. Indeed, the miPSC(þ) line 1, cultured under LIF(þ)/4SMs(þ) and LIF(À)/4SMs(À) conditions, did not significantly change during 24 h. Furthermore, the differences in cell shape between the two cultures did not appear to differ when analyzed with a human eye ( Figure 5A). The accuracy of the training and validation are close to 1, and the training and validation loss values are close to 0 in all time zones. The confusion matrix results also demonstrated a high accuracy to distinguish cells cultured in each medium in the entire time zones ( Figure 5B). Precision and recall values in both the LIF(þ)/4SMs(þ) and LIF(À)/4SMs(À) graphs decreased from 12 h due to cell proliferation and cell observation limitations caused by cell apoptosis ( Figure 5C). However, both values reached more than 0.9 points. Through this training result, it was determined that CNN can distinguish subtle differences in cell morphology with high accuracy. Subsequently, further experiments were conducted under the LIF(þ)/4SMs(þ) and ITS conditions. Since ITS induces cell differentiation, any morphological changes identified between the www.advancedsciencenews.com www.advintellsyst.com two culture conditions were relatively distinguishable by eye ( Figure S4a, Supporting Information). Therefore, CNN was able to analyze these with high accuracy ( Figure S4b, Supporting Information). The other miPSC cell line, miPSC(þ) line 2, was performed exactly as the miPSC(þ) line 1 training method (Table S1, Supporting Information). In the LIF(þ)/4SMs(þ) and LIF(À)/ 4SMs(À) groups, the accuracy of training and test was high in all time zones, but the test accuracy was only 64% at 24 h. The reason for the low accuracy was that the cells grew very rapidly over 24 h and accordingly more cell death occurred, making it difficult to observe the cells. In the LIF þ 4SMs and ITS groups, www.advancedsciencenews.com www.advintellsyst.com training and validation accuracies were close to 1; however, the test accuracy was lowest at 12 h. The CNN training results on miPSC(þ) line 2 can be considered a successful training results because the average accuracy was over 90%, although the accuracy was relatively lower than that of line 1. These results confirm that CNN can be applied to the various miPSC lines. In our previous study, the addition of the 4SMs to the cells reduces DNA damage and preserves high-quality iPSCs in long-term cultures. [29] However, since CNNs have been trained on iPSCs cultured under various cell culture conditions, the stem cell lines with genetic defects caused by double strand break can be distinguished during long-term culturing with cell images only by trained CNNs, without the need to identify them by immunocytochemistry and quantitative real-time PCR (qRT-PCR), etc. Ultimately, this can lead to copious amounts of time being saved spent in the experiments. www.advancedsciencenews.com www.advintellsyst.com

Prediction of the Morphological Changes in IPSCs with Genetic Defects
The miPSC(À) used in training was cultured with basic ESC media without the 4 SMs (P15). This means that miPSC(À) has many genetic defects, which promote numerous cell deaths during culturing. Indeed, the difference in cell shape according to the presence or absence of LIF was not significant when observed with the naked eye ( Figure 6A). Moreover, the cell morphology does not change significantly over 24 h. When trained with CNN, it was difficult to extract cell features due to cell apoptosis. In contrast to cell lines currently trained with CNN, the CNN accuracy results trained with miPSC(À) were not relatively high ( Figure 6B). The validation loss value is unstable at 1 and 24 h, and the accuracy of distinguishing between the two media conditions was the lowest at 77% ( Figure 6B, right table). Quantitative analysis of how much cell shape differs according to the two media conditions in all miPSC lines shows that there are no statistical differences in all time zones ( Figure S2A,B, Supporting Information). This means that it is impossible to analyze the difference in cell shape from the actual human eye, but using CNN technology, the subtle difference can be analyzed and distinguished. The ITS medium induced cell differentiation, which adequately changed the cell shape ( Figure S5A, Supporting Information). The difference between the two media conditions was well distinguished using the human eye, however, more cell apoptosis occurred following cultivation. Following 12 h of incubation, the cell's size became too large, and cell death obscured the observation of the cell's shape. Therefore, the validation loss at 12 and 24 h was overfitted and the accuracy of the test was not relatively low ( Figure S5B, Supporting Information). To overcome overfitting, the number of training images should be increased for certain cell lines, and the number of seeding cells should be reduced to obtain high-quality cell images.
Although the accuracy of miPSC(À) is relatively lower than that of other cells, it is still high enough to allow for the practical application of CNN technology. For this reason, we have confirmed that CNN could analyze the images of miPSC(þ) and miPSC(À) and distinguish from which culture medium condition the cells were cultured. Moreover, miPSC(þ) increased genetic stability without DNA damage occurring from the small molecules. In addition, the genomic stability of iPSCs was maintained following periods of long-term incubation. In this experiment, it was observed that iPSC(þ) maintains a characteristic round shape of iPSCs during cell culture. Conversely, since miPSC(À) is cultured without small molecules, DNA damage occurs during cell culture and in vitro differentiation, which reduces the properties and potential of the iPSCs and results in changes to the cell shape. Although it is difficult to visually distinguish miPSC(þ) and miPSC(À), the trained CNN can distinguish between the two cells. These results indicate that CNN may easily identify the quality of iPSCs, instead of performing assay experiments to identify genetic defects in cells, such as karyotypes, qRT-PCR, and immunostaining.

Prediction of the Genetic Defects in IPSCs
Next, we examined whether CNN could distinguish the cell morphological changes by culturing the cells in a culture condition containing the small molecules that have been known to affect genetic stability. Two cell lines were used: 1) the miPSC(þ) LT (long term) treated with 4SMs for 30 passages, 2) the miPSC(AE) ST (short term) treated with 4SMs for only three passages, and then incubated in their absence until passage 30. There are cell images obtained by culturing all miPSCs present in our laboratory for 6 h ( Figure 7A). Interestingly, the size and morphology of the cells were found to differ depending on the duration of treatment with the four SMs. The magnified cell images below show that both the miPSC(þ) line 1 and line 2 alongside the miPSC(þ) LT treated with the four SMs continuously exhibit small cell sizes and the characteristic round shape of the miPSCs. Conversely, these data show that as the four SMs are removed the cells grow larger and more widespread as the passage number increases. Therefore, training with CNN caused it to distinguish the culture medium from which the cells were cultured by only analyzing the image of miPSCs, with at least a 98% accuracy ( Figure 7B). The morphology and size of each cell are different depending on the duration of the treatment with the four SMs ( Figure 7C). A qRT-PCR experiment was conducted to determine the expression of ZSCAN4, indicating gene safety. Here, the longer the cells were treated with the four SMs, the higher the ZSCAN4 gene expression and the quality of the miPSCs ( Figure 7D). The treatment with the four SMs increased genetic safety, resulting in cell morphological differences. Due to the close relationship between the four SMs and DNA damageinduced genetic defects, CNN can distinguish culture conditions by only analyzing the cell images as opposed to conducting extensive biological experiments. Furthermore, it can also determine whether there are any genetic defects in the miPSCs.

Discussion
Currently, several papers have been published on efficient analysis methods using deep learning in the field of biology. For example, DNNs that detect auditory sensory cells and synapses, deep CNNs that automatically classify fluorescently labeled cells, and detect astrocytes with high accuracy, and CNNs that successfully analyze the morphology of mESCs and hiPSCs through culture media conditions. [28,[34][35][36] This study demonstrates that CNN can be trained using different cell images taken with transmission light microscopes over 24 h. In addition, CNN could correctly classify the morphological features of cells that are subtly altered by different culture medium conditions. The number of cell images used in the initial training was 1000, while 800 images were used in the validation set, and 100 images in the test set, all of which resulted in low accuracy and overfitting.
To solve these problems, the number of training images was increased to 2000, which consequently improved the training accuracy of most CNNs to nearly 100%. This proved that neural networks are very sensitive to morphological changes in cells.
Although this study appears to be similar to Waisman et al. [28] CNN method, however, there are several differences. The analysis accuracy was improved using more training images (2000 images) and the analytical ability of CNN was tested by applying more culture media conditions that caused cell morphology to change. Furthermore, Park et al. [29] discovered 4SMs that maintain the genetic stability of iPSCs, which result in subtle cellular morphological changes depending on the presence of 4SMs. As a result of training, we demonstrated CNN accurately distinguishes subtle cell morphological differences with the presence or absence of 4SMs with high accuracy.
There are several requirements that trained neural networks must meet to achieve high accuracy. First, the input data size was 240 Â 320 pixels, which started lower than the pixels used in other studies. This reduced computational effort allowed the CNN training to be completed quickly. Second, we created five DNNs and repeatedly constructed convolutional layers in them to create and train very deep CNNs (Supplementary Data). Third, to artificially increase the number of images provided to CNN, one image was cropped to 12 Â 12, and no image enhancement was performed. After meeting these conditions, we can check the process of how cell images are analyzed for each deep layer (Figure 4). The utilization of graphics processing units (GPUs) could be analyzed quickly based on a large amount of big data in a relatively short period, and these various requirements could Figure 7. Identifying the relationship between CNN and genetic defects. A) The miPSCs treated with 4SMs become round in shape, and miPSCs without 4SMs grow as the cell shape spreads. Considering the cell analysis method, the size of each cell is analyzed using Image J. Scale bars, 50 μm. B) All images are used as input data at 6 h after processing the medium for maintenance of each cell. The accuracy of both the training and validation are close to 1. On the right confusion matrix, CNN can distinguish all the cells with an accuracy of 98% or higher. C) This graph shows the cell size of the miPSCs. It is evident that the cell size varies alongside the 4SMs treatment duration, n = 6 biological repeats. D) Quantitative real-time PCR detection of the ZSCAN4 gene expression in all iPSC lines, n = 3 biological repeats. Data are shown as mean AE SEM. Statistical significance is determined by independent and paired Student's t-tests for unpaired and paired samples, respectively. *p < 0.05, **p < 0.01, ***p < 0.001.
www.advancedsciencenews.com www.advintellsyst.com be trained accurately and quickly by CNN in the next experimental study. All these factors will grow into an area where CNN relies on a lot of image analyses over the next few years. Consequently, most of the CNNs provided a training accuracy of almost 1 at 1 h after treatment with the culture medium and showed an average accuracy of more than 90% in the test accuracy. This occurred because of the large number of images being used for the training but because the number of hidden layers was increased. In addition, actual quantitative analysis demonstrates that there is no difference in cell morphology with the naked eye. However, CNN can analyze subtle differences and distinguish which culture medium conditions and which cell lines are. However, these also experienced limits. Indeed, the growth rate was different for each cell type, and the occurrence of apoptosis, in a particular cell, limited the cell morphology analysis and affected the overall accuracy. In addition, the colony size was too large to be fully captured in one image 24 h after the culture medium treatment (Figure 2-6). This produced limitations in distinguishing CNNs and resulted in relatively low accuracy. To overcome these limitations, the number of cells will be seeded differently according to the characteristics of each cell regardless of the cell type, and the culture medium should be frequently replaced.
We discovered that the shorter the processing period of 4SMs, the larger the size of the miPSCs ( Figure 7C). It can be inferred that genetic deficiency exists when the size of miPSCs spread and grow. RT-qPCR was conducted in biological experiments to accurately analyze the expression of genetic safety in cells ( Figure 7D). An analysis of ZSCAN4 gene expression, which shows chromosomal stability, confirmed that cells treated with 4SMs for a long time were genetically safe. However, as a result of the expression of the ZSCAN4 gene, miPSC(À) is genetically safer than miPSC(AE)ST. The reason for the low expression of miPSC(AE) ST can be thought of as a rapid progression of genetic deficiency due to rapid environmental changes by long-term removal of 4SMs after treatment of them. The difference in expression between miPSC(À) and miPSC(AE)ST is about twice as much, but the difference between miPSC(þ) LT and miPSC(À) is about 4 times, and miPSC(þ)LT and miPSC(AE)ST is about 10 times or more. This shows that miPSC(À) and miPSC(AE)ST can be grouped into cells with the same genetic defect. miPSCs with genetic safety and deficiency can be grouped by CNN analyzing only cell images.
In conclusion, we demonstrated that, in our iPSC and ESC models, CNN can distinguish the subtle changes in the cell morphologies that could be occurred between the presence or absence of the small molecules that prevent abnormal gene expression patterns by looking only at cell images without the need for labeling. CNN technique using image-based analysis can distinguish genetically safe high-quality stem cells and apply it to address the various challenges related to stem cell therapy and their clinical applications as advancing the stable maintenance and reproducible differentiation of the stem cells. Finally, the development of a robust system that detects highly precise and effective cell morphological changes in CNNs will soon surpass human capabilities of detection.
Cell Medium Conditions Applied to CNN Training: The mESCs and miPSCs were cultured with three different medium culture conditions, respectively, as follows: In mESCs, maintenance medium cell culture with the LIF, LIF-free culture medium, and DMEM/F12 (Gibco) (1:1) supplemented with 100 μg/mL Penicillin and Streptomycin (Gibco), ITS to differentiate mESCs into NP cells. In miPSC line 1(þ) and line 2(þ), maintenance medium cell culture containing four small molecules (SB431542, PD0325901, thiazovivin, AA, and LIF), maintenance medium cell culture without LIF and four small molecules, and DMEM/F12 supplemented with 100 μg/mL Penicillin and Streptomycin, ITS to induce differentiation. mESCs and miPSCs were incubated for 24 h for each culture medium condition to collect CNN training data.
Cell Image Collection and Image Processing: For the collection of the image data, stem cells cultured in each medium condition were taken with transmitted light microscopy (Olympus, Tokyo, Japan) at 1, 3, 6, 12, and 24 h. A total of 1 Â 10 5 cells were plated in 12-well plates (Corning) approximately 24 h before imaging. The 10Â objective was used with the light transmission. Image files were saved in the TIFF format. We randomly took images for each culture medium condition and conducted three biological replicates. The image pixels from microscopy were 2560 Â 1922 in three channels (RGB). Each picture was then resized to obtain images of 480 Â 640 pixels by applying the python script ImageSlicer. Approximately 400 cell images were taken in each group, and one image was split into 12 using the Python "image slicer" function to increase the number of images being used for the CNN training. In total, approximately 4800 images were obtained, and the network contained 2000 images for training in each group alongside 400 images for validation in each group. A further 100 images per group were reserved for the test after training. To apply the CNN`s ResNet-50 algorithm, image pixels were downsized to 240 Â 320 for the training. Independent replicates (n = 3) were processed in the same way.
Purification of Total RNA and qRT-PCR: Total RNA was prepared using TRI reagent (MRS, Cincinnati, OH), following the manufacturer's instructions. The RNA content was quantified by a Nanodrop One microvolume UV-Vis spectrophotometer (Thermofisher). The cDNA was synthesized using 5 μg of total RNA using M-MLV reverse transcriptase (Thermofisher) and oligo(dT) primers at 42°C for 1 h. SYBR green-based qRT-PCR was performed in triplicate with KAPA SYBR FAST ABI prism qPCR kit reagents (KAPA Biosystems) and the relative quantification was determined using the StepOnePlus Real-Time PCR system (Life Technologies Holdings Pte Ltd). All samples were then calculated by means of the comparative Ct method (2-ΔΔCt) relative to the expression of the glyceraldhyde-3-phosphate dehydrogenase control. PCR conditions were: 10 min at 95°C, followed by 50 cycles of 20 s at 95°C, 20 s at 60°C, and 20 s at 72°C, finally followed by 5 min at 72°C. The promoter region of the ZSCAN4 gene was amplified using the following primers (F: 5 0 -GAG ATT CAT GGA GAG TCT GAC TGA GTG-3 0 ; R: 5 0 -GCT GTT TCA AAA GCT TGA CTT-3 0 ). www.advancedsciencenews.com www.advintellsyst.com