Method for identifying pluripotency of stem cells based on Raman spectrum and generative adversarial network

Abstract

The invention discloses a method for identifying pluripotency of stem cells based on a Raman spectrum and a generative adversarial network. The method comprises the following steps: acquiring a data set, culturing iPSC cells in a culture medium, collecting spectrums of the iPSC cells in the culture process by using a Raman spectrometer, collecting data of a plurality of time periods, grouping according to the time periods, and collecting SERS signals of a plurality of cells in each group; a data set is constructed, in a time period, according to the sequence from front to back, six stages are taken, Raman spectrums of cell samples in each stage are divided into one class, and a spectrum data set can be obtained and comprises six classes of spectrums; data preprocessing: performing baseline removal, normalization and other processing on the data set; training a generative adversarial network model; testing the generative adversarial network model; detecting the pluripotency of the iPSC; and searching a key spectral region for judging cell pluripotency. According to the method, the pluripotency of the stem cells can be recognized in a lossless and high-precision manner, and the corresponding biomarkers can be screened.

Description

Technical Field

[0001] The invention relates to a stem cell identification technology, in particular to a method for identifying stem cell pluripotency based on Raman spectroscopy and a generative adversarial network. Background Art

[0002] Induced pluripotent stem cells (iPSCs) have broad application prospects in regenerative medicine, drug screening, and disease model construction due to their potential for self-renewal and differentiation into a variety of cell types. The pluripotency of iPSCs is the key to their application in regenerative medicine, disease modeling, and drug screening. This ability enables iPSCs to differentiate into a variety of different cell types, providing a wide range of application possibilities and great research value. Currently, methods for detecting iPSC pluripotency mainly rely on genomic, proteomics, and metabolomics techniques, such as immunocytochemistry, flow cytometry, Western blot, and polymerase chain reaction. Although these methods are effective, they are usually complex, time-consuming, and destructive, and may damage the integrity of stem cells. Moreover, these methods require the knowledge of specific biomarkers in advance, which further affects their clinical application.

[0003] In recent years, Raman spectroscopy has gradually become a promising technology in stem cell research due to its high sensitivity, label-free and non-destructive detection characteristics. By analyzing the molecular vibration modes in cells, Raman spectroscopy can provide rich chemical and structural information, especially in the real-time characterization of iPSC pluripotency at the single-cell level, which has unique advantages. Unlike traditional analysis methods, Raman spectroscopy does not rely on predetermined markers or external labels, so it can directly analyze the internal state of stem cells in a more simplified way. Especially in the characterization of iPSC pluripotency and its dynamic changes, Raman spectroscopy shows broad application potential.

[0004] However, the large amount of high-dimensional data generated by Raman spectroscopy often requires complex computational processing, and traditional analysis methods such as principal component analysis (PCA) and support vector machine (SVM) have certain limitations in processing these data, especially in terms of automation and classification accuracy. Therefore, it is of great significance to develop a technology that does not rely on tedious manual feature selection and can non-destructively and accurately evaluate the pluripotency of stem cells, and can screen the corresponding biomarkers. Summary of the invention

[0005] In view of the above-mentioned defects, the purpose of the present invention is to propose a method for stem cell pluripotency identification based on Raman spectroscopy and generative adversarial network (GAN), which can identify stem cell pluripotency non-destructively and with high precision.

[0006] To achieve this object, the present invention adopts the following technical solutions:

[0007] like Fig. 9 As shown, the method for identifying stem cell pluripotency based on Raman spectroscopy and generative adversarial network includes the following steps:

[0008] S100, data set acquisition, iPSC cells are cultured in culture medium, and the spectra of iPSC cells during the culture process are collected using a Raman spectrometer. Data of multiple time periods are collected and grouped by time period, and SERS signals of multiple cells are collected in each group;

[0009] S200, constructing a data set. Within a time period, six stages are taken in order from the front to the back. The Raman spectra of the cell samples in each stage are divided into one category. The spectral data set can be obtained, which contains six categories of spectra.

[0010] S300, data preprocessing, using WiRE 4.3 software to remove baseline, remove cosmic rays, reduce noise and normalize the obtained Raman spectrum data, and randomly extract the processed spectrum data into training set, validation set and test set in a ratio of 7:2:1;

[0011] S400, training of a generative adversarial network model. The generative adversarial network model consists of two parts: a generator and a discriminator. The generator is designed to generate fake data similar to the input spectral data from random noise. The main function of the discriminator is to distinguish the data generated by the generator from the real spectral data. During the training process, the discriminator adopts a convolutional neural network architecture to enhance the feature extraction capability.

[0012] S500, testing the generative adversarial network model, preserving the test set that has not participated in the network training in the data set in step S300, and directly inputting the spectrum into the trained model to quickly determine the category to which the spectrum belongs;

[0013] S600, detect the pluripotency of iPSC: judge the current state of cells based on the model prediction results of the generative adversarial network, and quickly understand the pluripotency of current cells.

[0014] Preferably, the method further includes step S700, which is to find the key spectral region for judging the pluripotency of cells: the discriminator of the generative adversarial network is additionally added with the forward and backward hook functions of Grad-CAM to capture the gradient information and feature map of the convolution layer, calculate the weight corresponding to each category through back propagation, and multiply the weight by the feature map to generate the final CAM heat map, thereby intuitively showing the key spectral region of the discriminator in each category of decision. The same is applicable to finding the key spectral intervals of other cells.

[0015] Further, in step S100, iPSC cells are cultured with DMEM medium, and cells are taken at 0, 1, 3, 6, 12, and 24 hours during the culture process, and SERS spectra of the cells at these 6 stages are collected using a Raman spectrometer;

[0016] The laser wavelength was 633 nm, the laser power was set to 50%, and the L 50× objective lens with a numerical aperture of 0.5 was used to collect 400 to 1800 cm -1 Raman spectra in a range of wavenumbers;

[0017] For each group of samples, at least 50 cells were selected for SERS signal acquisition, and each cell sample was measured at least 15 times.

[0018] Furthermore, in step S200, the pluripotency of iPSC is divided into 6 stages according to the culture time of DMEM medium, and the cells and their Raman spectra are correspondingly divided into 6 categories; the Raman spectra of the cell samples in each stage are divided into one category, and a spectral data set is obtained, including 6 categories of spectra of iPSC initial time, 1h, 3h, 6h, 12h and 24h.

[0019] Further, in step S300, the preprocessed spectral data set contains at least 6000 Raman spectra of single cells.

[0020] Furthermore, in step S400, the generator aims to generate forged data similar to the input spectral data from random noise, and its input layer is 100-dimensional random noise. After three fully connected layers, the output dimensions are 256, 512 and 1242 respectively, ensuring that the dimensions of the generated forged data are consistent with those of the real spectral data; the ReLU activation function is applied after each fully connected layer to introduce nonlinear characteristics, thereby enhancing the expression ability of the model; GAN uses adversarial training of the generator and the discriminator to enable the generator to gradually generate more realistic spectral data;

[0021] The main function of the discriminator is to distinguish the data generated by the generator from the real spectral data. Its input is 1242-dimensional spectral data. During the training process, the discriminator adopts a convolutional neural network (CNN) architecture to enhance the feature extraction capability. The network contains four convolutional layers and four maximum pooling layers. The number of convolutions is 16, 32, 64, and 128, respectively. The convolution kernel size is unified as 1×3. The activation function is ReLU. The window size of the pooling layer is 1×3, and the step size is 1. It then passes through three fully connected layers, which are 256, 128 dimensions, and the output 6-dimensional multi-classification layer. Each layer is followed by a ReLU activation function and a Dropout layer to prevent overfitting and ensure the generalization ability of the model, which is suitable for multi-classification tasks.

[0022] The Adam optimizer is used in the model training process, and the learning rate is set to 8×10-4 , the loss function is cross entropy loss, and the batch size is 36; to further optimize the training, the model uses the StepLR learning rate scheduler, which is set to step_size=30 and gamma=0.1 for the generator and discriminator respectively.

[0023] One of the above technical solutions includes the following beneficial effects: the method for rapid, non-destructive and high-precision detection of stem cell pluripotency based on Raman spectroscopy and generative adversarial network GAN proposed by the present invention does not require the destructive and lengthy sample pretreatment process of the traditional detection method, nor does it require prior knowledge of the corresponding biomarkers. The biological molecular information of different pluripotent stem cells is obtained through Raman spectroscopy, and the detection of different pluripotencies of stem cells can be achieved by training the generative adversarial network GAN. Therefore, the present invention can achieve rapid, non-destructive and high-precision detection of stem cell pluripotency. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is the induction scheme of induced pluripotent stem cells, phase contrast images of different pluripotencies, immunofluorescence images and fluorescence intensity histograms;

[0025] Figure 2 Single cell average Raman spectra of six cell states during stem cell culture in DMEM medium;

[0026] Figure 3 for Figure 2 Comparison of peak signal intensity of characteristic peaks;

[0027] Figure 4 The loss function for training the generative adversarial network 200 times;

[0028] Figure 5 Training accuracy for 200 training runs of the generative adversarial network;

[0029] Figure 6 To generate the ROC curve for each category of the adversarial network model;

[0030] Figure 7 To generate the confusion matrix of the test set during adversarial network training;

[0031] Figure 8 Result plots of key spectral regions generated for Grad-CAM;

[0032] Fig. 9 It is a schematic diagram of the method flow of this scheme. DETAILED DESCRIPTION

[0033] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0034] The method for identifying stem cell pluripotency based on Raman spectroscopy and generative adversarial network includes the following steps:

[0035] S100, data set acquisition, iPSC cells are cultured in culture medium, and the spectra of iPSC cells during the culture process are collected using a Raman spectrometer. Data of multiple time periods are collected and grouped by time period, and SERS signals of multiple cells are collected in each group;

[0036] S200, constructing a data set. Within a time period, six stages are taken in order from the front to the back. The Raman spectra of the cell samples in each stage are divided into one category. The spectral data set can be obtained, which contains six categories of spectra.

[0037] S300, data preprocessing, using WiRE 4.3 software to remove baseline, remove cosmic rays, reduce noise and normalize the obtained Raman spectrum data, and randomly extract the processed spectrum data into training set, validation set and test set in a ratio of 7:2:1;

[0038] WiRE is based on WiRE is the Raman Environment software for Renishaw's Raman instruments. WiRE controls Raman data acquisition and provides users with specialist data processing and analysis options including baseline, cosmic ray removal, noise reduction and normalisation. With WiRE users can identify unknown spectra, remove background interference or determine the distribution of particles in a megapixel Raman image.

[0039] S400, training of a generative adversarial network model. The generative adversarial network model consists of two parts: a generator and a discriminator. The generator is designed to generate fake data similar to the input spectral data from random noise. The main function of the discriminator is to distinguish the data generated by the generator from the real spectral data. During the training process, the discriminator adopts a convolutional neural network architecture to enhance the feature extraction capability.

[0040] S500, testing the generative adversarial network model, preserving the test set that has not participated in the network training in the data set in step S300, and directly inputting the spectrum into the trained model to quickly determine the category to which the spectrum belongs;

[0041] S600, detect the pluripotency of iPSC: judge the current state of cells based on the model prediction results of the generative adversarial network, and quickly understand the pluripotency of current cells.

[0042] The present invention proposes a method for rapid, non-destructive, and high-precision detection of stem cell pluripotency based on Raman spectroscopy and generative adversarial network GAN. Compared with the traditional method for detecting cell status based on genomics, proteomics, and metabolomics, the present invention does not require the destructive and lengthy sample pretreatment process of the traditional detection method, nor does it require prior knowledge of the corresponding biomarkers. The biological molecular information of different pluripotent stem cells is obtained through Raman spectroscopy, and the detection of different pluripotencies of stem cells can be achieved by training the generative adversarial network GAN. Therefore, the present invention can achieve rapid, non-destructive, and high-precision detection of stem cell pluripotency.

[0043] Compared with the traditional CNN model, GAN can not only learn classification features, but also expand the spectral data set, generate pseudo data similar to real spectral features, and enhance data diversity. It is especially suitable for situations where spectral data is scarce, which helps to improve the robustness of the model. The main function of the discriminator is to distinguish between the data generated by the generator and the real spectral data. Its input is 1242-dimensional spectral data.

[0044] like Figure 8 As shown, it also includes step S700, finding the key spectral region for judging cell pluripotency: the discriminator of the generative adversarial network adds the forward and backward hook functions of Grad-CAM to capture the gradient information and feature map of the convolution layer, calculate the weight corresponding to each category through back propagation, and multiply the weight with the feature map to generate the final CAM heat map, so as to intuitively show the key spectral region of the discriminator in each category of decision. It is also applicable to finding the key spectral intervals of other cells.

[0045] Through deep learning, Raman spectra are identified, different spectral regions are predicted independently, and the classification features of the generative adversarial network are visualized using gradient-weighted class activation maps to accurately identify and distinguish the pluripotency of stem cells and screen potential biomarkers. Therefore, the combination of Raman spectroscopy technology and deep learning helps to achieve rapid, non-destructive and accurate detection of stem cell pluripotency, and automatically screen stem cell biomarkers through label-free methods, thereby improving the safety and effectiveness of its clinical application.

[0046] like Figure 1-3 As shown, in step S100, iPSC cells are cultured with DMEM medium, and cells at 0, 1, 3, 6, 12, and 24 hours are taken during the culture process, and SERS spectra of the cells at these 6 stages are collected using a Raman spectrometer;

[0047] The laser wavelength was 633 nm, the laser power was set to 50%, and the L 50× objective lens with a numerical aperture of 0.5 was used to collect 400 to 1800 cm -1Raman spectra in a range of wavenumbers;

[0048] For each group of samples, at least 50 cells were selected for SERS signal acquisition, and each cell sample was measured at least 15 times.

[0049] This experiment uses 633nm wavelength laser for excitation, which can effectively reduce photothermal effects and fluorescence interference compared to short-wavelength lasers, and is suitable for biological sample detection. The laser power is set to 50%, which ensures signal intensity while avoiding sample damage; combined with a 50× objective lens with a numerical aperture of 0.5, efficient signal acquisition and high-resolution detection of local areas within cells are achieved. The acquisition wavenumber range is set to 400-1800cm-1, covering the main characteristic Raman peaks of biological molecules and providing rich chemical information. At least 50 cells were collected for each group of samples, and each cell was measured 15 times, which can not only improve the statistical significance of the data, but also reduce random noise by signal averaging. The design of these parameters fully balances signal quality, sample protection, and data reliability, providing a solid foundation for Raman spectroscopy analysis at the cellular level. Figure 3 By comparing the peak signal intensities of characteristic peaks, the differences in molecular characteristics of cells at different pluripotency stages were revealed, providing a physical basis for the classification of the GAN model.

[0050] Among them, in step S200, the pluripotency of iPSC is divided into 6 stages according to the culture time of DMEM culture medium, and the cells and their Raman spectra are correspondingly divided into 6 categories; the Raman spectra of cell samples in each stage are divided into one category, and a spectral data set is obtained, including 6 categories of spectra of iPSC initial time, 1h, 3h, 6h, 12h and 24h.

[0051] The pluripotency state of iPSC is divided into six stages according to the DMEM culture time: initial time, 1h, 3h, 6h, 12h and 24h, and the cells and their Raman spectra are classified, which helps to capture the dynamic changes of pluripotency. Early short time intervals (such as 1h, 3h, 6h) can reflect rapidly changing molecular characteristics, while later long intervals (12h, 24h) are suitable for recording changes in state tendencies, which not only improves the scientific nature of the data but also reduces redundancy. Through this time division, a balanced spectral data set can be generated, providing a clear training target for the GAN model and improving its recognition accuracy. At the same time, this division scheme supports the real-time monitoring needs of the pluripotency state and provides strong support for dynamic evaluation and quality control during stem cell culture.

[0052] Wherein, in step S300, the preprocessed spectral data set contains at least 6000 Raman spectra of single cells.

[0053] Using at least 6,000 single-cell Raman spectra as a dataset provides a wealth of training samples for the GAN model, enhances the generalization ability and classification accuracy of the model, and captures the heterogeneity characteristics between single cells. This scale of dataset not only meets the sample size requirements of deep learning, but also improves the statistical reliability of the analysis and the robustness of the spectral signal, providing a solid data foundation for the study of stem cell pluripotency identification.

[0054] In addition, in step S400, the generator aims to generate forged data similar to the input spectral data from random noise, and its input layer is 100-dimensional random noise. After three fully connected layers, the output dimensions are 256, 512 and 1242 respectively, ensuring that the generated forged data is consistent with the dimension of the real spectral data; the ReLU activation function is applied after each fully connected layer to introduce nonlinear characteristics, thereby enhancing the expressive power of the model; GAN enables the generator to gradually generate more realistic spectral data through adversarial training of the generator and the discriminator; compared with the traditional CNN model, GAN can not only learn classification features, but also expand the spectral data set to generate pseudo data similar to the real spectral features, enhance data diversity, and is particularly suitable for situations where spectral data is scarce, which helps to improve the robustness of the model.

[0055] The main function of the discriminator is to distinguish the data generated by the generator from the real spectral data. Its input is 1242-dimensional spectral data. During the training process, the discriminator adopts a convolutional neural network (CNN) architecture to enhance the feature extraction capability. The network contains four convolutional layers and four maximum pooling layers. The number of convolutions is 16, 32, 64, and 128, respectively. The convolution kernel size is unified as 1×3. The activation function is ReLU. The window size of the pooling layer is 1×3, and the step size is 1. Then it passes through three fully connected layers, which are 256, 128 dimensions, and the output 6-dimensional multi-classification layer. Each layer is followed by a ReLU activation function and a Dropout layer to prevent overfitting and ensure the generalization ability of the model, which is suitable for multi-classification tasks. Compared with the ordinary CNN model, the discriminator can more keenly capture the local features of spectral data through the recognition of small differences in adversarial training, thereby improving the classification performance and generalization ability.

[0056] The Adam optimizer is used in the model training process, and the learning rate is set to 8×10 -4 , the loss function is cross entropy loss, and the batch size is 36; to further optimize the training, the model uses the StepLR learning rate scheduler, with step_size = 30 and gamma = 0.1 for the generator and discriminator respectively. This means that after every 30 epochs of training, the learning rate will be multiplied by 0.1, thereby achieving gradual decay, which helps to gradually slow down the learning rate in the later stages of training and stabilize the weight update.

[0057] This optimization strategy improves training efficiency and model performance, allowing the model to adapt to parameter update requirements at different stages. Through adversarial training, the spectral data generated by the GAN model is difficult to distinguish from real data, thereby improving the model's generation ability and robustness. In the PyTorch framework, after multiple training of the training set and validation set, the model with the highest validation accuracy is finally saved. Figure 4 It shows that the loss function converges smoothly after 200 trainings of the generative adversarial network, proving that the model training is stable and the generator and discriminator reach an optimal balance; Figure 5 The model training accuracy gradually improved and stabilized, indicating that GAN can effectively capture the characteristics of spectral data and has strong classification and generalization capabilities; Figure 6 To generate the ROC curve of each category of the adversarial network model, the AUC was greater than 0.94. The closer the ROC curve was to 1, the better the recognition accuracy was. These results jointly demonstrated the reliability and effectiveness of the model in identifying the pluripotency of stem cells.

[0058] The technical principle of the present invention is described above in conjunction with specific embodiments. These descriptions are only for explaining the principle of the present invention and cannot be interpreted as limiting the scope of protection of the present invention in any way. Based on the explanations herein, those skilled in the art can associate other specific implementations of the present invention without paying creative labor, and these methods will fall within the scope of protection of the present invention.

Claims

1. A method for identifying stem cell pluripotency based on Raman spectroscopy and generative adversarial networks, characterized in that: The steps include: S100, data set acquisition, iPSC cells are cultured in culture medium, and the spectra of iPSC cells during the culture process are collected using a Raman spectrometer. Data of multiple time periods are collected and grouped by time period, and SERS signals of multiple cells are collected in each group; S200, constructing a data set. Within a time period, six stages are taken in order from the front to the back. The Raman spectra of the cell samples in each stage are divided into one category. The spectral data set can be obtained, which contains six categories of spectra. S300, data preprocessing, using software to remove baseline, remove cosmic rays, reduce noise and normalize the obtained Raman spectrum data, and randomly extract the processed spectrum data into training set, validation set and test set in a ratio of 7:2:1; S400, training of a generative adversarial network model. The generative adversarial network model consists of two parts: a generator and a discriminator. The generator is designed to generate fake data similar to the input spectral data from random noise. The main function of the discriminator is to distinguish the data generated by the generator from the real spectral data. During the training process, the discriminator adopts a convolutional neural network architecture to enhance the feature extraction capability. S500, testing the generative adversarial network model, preserving the test set that has not participated in the network training in the data set in step S300, and directly inputting the spectrum into the trained model to quickly determine the category to which the spectrum belongs; S600, detect the pluripotency of iPSC: judge the current state of cells based on the model prediction results of the generative adversarial network, and quickly understand the pluripotency of current cells.

2. The method for identifying stem cell pluripotency based on Raman spectroscopy and generative adversarial networks according to claim 1, characterized in that: The method also includes step S700, which is to find the key spectral region for judging the pluripotency of cells: the discriminator of the generative adversarial network is additionally added with the forward and backward hook functions of Grad-CAM to capture the gradient information and feature map of the convolution layer, calculate the weight corresponding to each category through back propagation, and multiply the weight with the feature map to generate the final CAM heat map, thereby intuitively showing the key spectral region of the discriminator in each category of decision. It is also applicable to finding the key spectral intervals of other cells.

3. The method for identifying stem cell pluripotency based on Raman spectroscopy and generative adversarial networks according to claim 2, characterized in that: In step S100, iPSC cells are cultured in DMEM medium, and cells at 0, 1, 3, 6, 12, and 24 hours are taken during the culture process, and SERS spectra of the cells at these six stages are collected using a Raman spectrometer; The laser wavelength was 633 nm, the laser power was set to 50%, and the L 50× objective lens with a numerical aperture of 0.5 was used to collect 400 to 1800 cm -1 Raman spectra in a range of wavenumbers; For each group of samples, at least 50 cells were selected for SERS signal acquisition, and each cell sample was measured at least 15 times.

4. The method for identifying stem cell pluripotency based on Raman spectroscopy and generative adversarial networks according to claim 3, characterized in that: In step S200, the pluripotency of iPSC is divided into 6 stages according to the culture time of DMEM medium, and the cells and their Raman spectra are divided into 6 categories accordingly; the Raman spectra of the cell samples in each stage are divided into one category, and a spectral data set is obtained, including 6 categories of spectra of iPSC initial time, 1h, 3h, 6h, 12h and 24h.

5. The method for identifying stem cell pluripotency based on Raman spectroscopy and generative adversarial networks according to claim 4, characterized in that: In step S300, the preprocessed spectral data set contains at least 6000 Raman spectra of single cells.

6. The method for identifying stem cell pluripotency based on Raman spectroscopy and generative adversarial networks according to claim 5, characterized in that: In step S400, the generator aims to generate fake data similar to the input spectral data from random noise. Its input layer is 100-dimensional random noise. After three fully connected layers, the output dimensions are 256, 512 and 1242 respectively, ensuring that the dimensions of the generated fake data are consistent with the dimensions of the real spectral data. The ReLU activation function is applied after each fully connected layer to introduce nonlinear characteristics, thereby enhancing the expression ability of the model. GAN uses adversarial training between the generator and the discriminator to enable the generator to gradually generate more realistic spectral data. The main function of the discriminator is to distinguish the data generated by the generator from the real spectral data. Its input is 1242-dimensional spectral data. During the training process, the discriminator adopts a convolutional neural network (CNN) architecture to enhance the feature extraction capability. The network contains four convolutional layers and four maximum pooling layers. The number of convolutions is 16, 32, 64, and 128, respectively. The convolution kernel size is unified as 1×3. The activation function is ReLU. The window size of the pooling layer is 1×3, and the step size is 1. It then passes through three fully connected layers, which are 256, 128 dimensions, and the output 6-dimensional multi-classification layer. Each layer is followed by a ReLU activation function and a Dropout layer to prevent overfitting and ensure the generalization ability of the model, which is suitable for multi-classification tasks. The Adam optimizer is used in the model training process, and the learning rate is set to 8×10 -4 , the loss function is cross entropy loss, and the batch size is 36; to further optimize the training, the model uses the StepLR learning rate scheduler, which is set to step_size=30 and gamma=0.1 for the generator and discriminator respectively.

Images