Raman spectrum and CNN-LSTM network regression model integration-based iPSC pluripotency regression prediction method

Abstract

The invention belongs to the technical field of stem cell recognition, and particularly relates to an iPSC pluripotency regression prediction method based on integration of a Raman spectrum and a CNN-LSTM network regression model, and the method comprises the following steps: obtaining a data set, culturing iPSC in a culture medium, collecting the Raman spectrum of the iPSC in the culture process by using a Raman spectrometer, collecting the Raman spectrums of a plurality of cells in each time period, and calculating the pluripotency of the iPSC. Grouping is carried out according to time periods; carrying out pretreatment on the collected Raman spectrum; constructing and dividing a data set, dividing the Raman spectrums of the cell samples into six classes according to a time sequence, and constructing and randomly dividing the data set; the CNN-LSTM model is trained and tested, and the pluripotency of the iPSC is evaluated; through a regression method, the discrete pluripotent state and the transition capability gradient of the stem cells are accurately and quantitatively evaluated.

Description

Technical Field

[0001] The present invention belongs to the technical field of stem cell identification, and particularly relates to a regression prediction method for iPSC pluripotency based on the integration of Raman spectroscopy and a CNN-LSTM network regression model. Background Art

[0002] Induced pluripotent stem cells (iPSCs) have important scientific significance and application value in regenerative medicine, human disease modeling, tissue engineering, etc. due to their potential for self-renewal and differentiation into various cell types. They can promote innovation in medicine, drug development, and regenerative medicine.

[0003] The pluripotency of iPSCs (iPSCs) allows them to differentiate into a variety of cell types, such as cardiomyocytes, neurons, and pancreatic islet cells, enabling their potential use in treating a wide range of diseases. These cells have broad applications in disease modeling, personalized medicine, developmental biology, and basic research. Their unique pluripotency and plasticity make them essential tools in modern biomedical research and clinical applications. The clinical translation of iPSC-derived therapies relies heavily on rigorous quality control of pluripotency, a dynamic cellular state that requires continuous monitoring and precise assessment prior to differentiation to ensure cell consistency and therapeutic safety. Currently, methods for assessing iPSC pluripotency primarily rely on genomic, proteomic, and metabolomic techniques, such as immunocytochemistry, flow cytometry, Western blot, and polymerase chain reaction. These methods are cumbersome, time-consuming, and uneconomical, as well as being destructive and invasive, potentially compromising stem cell integrity. Furthermore, these methods require the prior knowledge of specific biomarkers, further hindering their practical application.

[0004] Raman spectroscopy, an inelastic light scattering spectroscopic technique based on the Raman scattering effect, can non-destructively reveal the chemical composition and structure of a sample, enabling insights into the molecular dynamics of fundamental cellular processes. Raman spectroscopy contains valuable information about molecular bonds and structure, enabling the identification of chemical components within a substance. Consequently, the technique is widely used in areas such as molecular structure identification, industrial process control, and life science research. In the life sciences, Raman spectroscopy is particularly well-suited for stem cell research, enabling the non-destructive monitoring of molecular changes and the pluripotency state of stem cells, providing crucial technical support for stem cell culture, differentiation, and application. Unlike traditional analytical methods, Raman spectroscopy does not rely on pre-defined markers or external labels, enabling a more streamlined approach to directly analyze the internal state of stem cells, providing non-destructive, real-time molecular information. Raman spectroscopy demonstrates unique advantages in characterizing the pluripotency of iPSCs and their dynamic changes, providing a powerful tool for monitoring stem cell differentiation, assessing pluripotency, and studying cellular reprogramming processes, with broad potential for application.

[0005] However, existing spectral analysis strategies struggle to capture subtle changes in cell states during differentiation and are unable to accurately quantify the dynamic changes in stem cells as they transition from one pluripotent state to another. Furthermore, the change in cell state is a spatiotemporal dynamic process, requiring simultaneous consideration of information in both temporal and spatial dimensions to fully understand the pluripotency state. Traditional spectral analysis methods, such as principal component analysis, linear discriminant analysis, and partial least squares regression, often only provide static spectral information and struggle to effectively extract accurate spatiotemporal spectral patterns. Deep learning methods, such as convolutional neural networks (CNNs), gated recurrent units (GRUs), and long short-term memory (LSTM) networks, primarily apply discrete state classification to stem cell pluripotency analysis. However, this approach loses continuity between stem cell states, failing to reflect the dynamic transition process. Furthermore, it cannot quantify the gradient of ability that characterizes transitions between states, making it difficult to capture the gradual process of stem cell dynamic transformation. CNNs and LSTMs have limitations in processing sequential data and spatiotemporal dynamics. Therefore, it is of great significance to develop a technology that does not rely on tedious manual feature selection, can simultaneously process the spatial and temporal characteristics of spectral data, and non-destructively and highly accurately evaluate the gradient quantification of stem cell pluripotency and transition ability. Summary of the Invention

[0006] The purpose of the present invention is to provide a regression prediction method for iPSC pluripotency based on the integration of Raman spectroscopy and CNN-LSTM network regression model, through which the discrete pluripotency states and transition ability gradients of stem cells can be accurately quantitatively evaluated.

[0007] The technical solutions adopted by the present invention are as follows:

[0008] The iPSC pluripotency regression prediction method based on the integration of Raman spectroscopy and CNN-LSTM network regression model includes the following steps:

[0009] Step 1: Acquire a dataset by culturing iPSCs in culture medium and using a Raman spectrometer to collect Raman spectra of the iPSCs during the culture process. The acquisition process is divided into multiple time periods, and spectral data of multiple cells are collected in each time period.

[0010] Step 2: Data preprocessing: removing baseline, removing cosmic rays, reducing noise and normalizing the Raman spectrum data;

[0011] Step 3: Construct and divide the data set. Within a time period, the Raman spectra of the cell samples are divided into six stages in chronological order. The spectral data of each stage is regarded as one class, for a total of six spectral data sets. The processed spectral data are randomly divided into training set, validation set and test set.

[0012] Step 4: Build and train a CNN-LSTM network regression model. The CNN-LSTM network regression model consists of two parts: a CNN and an LSTM. The CNN layer is used to extract the key spectral features of the Raman fingerprint, and the LSTM layer learns the long-term dependencies of the time series, combining the spatial features of the Raman spectrum with the time series dependencies to identify spectra in different time periods.

[0013] Step 5: The trained CNN-LSTM network regression model is used to evaluate the pluripotency of iPSCs. The reserved test set from Step 3, which did not participate in network training, is directly fed into the trained model. This test set is not used in training or hyperparameter tuning, but is used for final model performance evaluation to predict the ability of stem cells to transition between different states. The network evaluates the pluripotency of stem cells, quantifying the pluripotency state as a continuous value, and the regression model predicts the specific value of each state.

[0014] Preferably, in step 1, during the iPSC culture, the iPSC culture medium is replaced with RPMI1640 culture medium at 1, 3, 6, 12, and 24 hours, respectively, to affect iPSC differentiation, and spectra of the iPSC cells at the five time points and the original iPSC cells without culture medium replacement are collected at a total of six time points;

[0015] The present invention collects data with a laser wavelength of 633 nm, a laser power of 50%, a L50× objective lens with a numerical aperture of 0.5, and collects data from 600-1800 cm -1 Raman spectra in a wide range of wavenumbers; for each sample at each time point, at least 50 cells were selected for signal acquisition, and each cell sample was measured at least 15 times.

[0016] Preferably, in step 1, the fluorescence intensity values ​​of stem cell markers SOX2 and OCT4 are detected by immunofluorescence staining; and the average immunofluorescence intensity of the iPSC state is calculated using ImageJ, where average fluorescence intensity = total fluorescence intensity of the colony region / area of ​​the colony region.

[0017] Preferably, in step 2, when preprocessing the obtained data, WiRE4.3 software is used, including baseline removal, cosmic ray removal, and noise reduction. Before the data set is input into the network for training, the spectral data is normalized and the data values ​​are scaled to the range of [0,1] to eliminate the potential impact of spectral intensity changes on the model results.

[0018] Preferably, in step 3, the Raman spectrum is divided into six stages according to the chronological order of processing, and the Raman spectrum data is randomly extracted into a training set, a validation set, and a test set in a ratio of 8:1:1; the random division process uses stratified sampling to ensure the balance of category distribution and to ensure that the spectral data of the same cell will only appear in one of the training set, validation set, and test set, thereby ensuring the reliability of the results; the divided training set, validation set, and test set are respectively used for model training, hyperparameter tuning, and performance evaluation.

[0019] Preferably, in step 4, the CNN-LSTM network regression model consists of a CNN layer and an LSTM layer; the local features of the spectral data are extracted by CNN, the input data dimension is the spectral length, and after four layers of convolution with convolution kernel sizes of 12, 14, 16 and 16, the output dimensions are 16, 32, 64 and 128 respectively; the LSTM layer performs time series modeling on the 128-dimensional feature sequence, captures the time series dependency through four layers of bidirectional LSTM, and then maps the features to the regression values ​​of 6 time points through a fully connected layer, corresponding to the pluripotency scores of the two markers SOX2 and OCT4, to achieve high-precision pluripotency assessment.

[0020] In step 4, during the training of the CNN-LSTM network regression model, Adam is used as the optimizer to train the model, and the learning rate of Adam is set to 1×10 -5 , and used the mean square error loss as the loss function; built a CNN-LSTM network regression model through PyTorch, and put the training set and validation set after the data set processing in step 3 into the CNN-LSTM network regression model for training. At the end of each round, the model performance was evaluated on the validation set, and the training loss (Train Loss), validation loss (Val Loss), root mean square error (Root Mean Square Error, RMSE) and determination coefficient R were recorded. 2 .

[0021] The technical effects achieved by the present invention are:

[0022] This invention enables quantitative assessment of pluripotency. An adaptive wide-kernel convolution module extracts key features of Raman spectra, combined with an LSTM time-dependent learning module to capture dynamic changes in spectral sequences, improving the accuracy of identifying spectral details and temporal features. The CNN-LSTM model, through regression prediction of Raman spectra, preserves continuity between stem cell states, accurately reflects the dynamic transition process of stem cells, and quantitatively characterizes the capacity gradient between state transitions, achieving accurate assessment of pluripotency and transitional capacity.

[0023] Compared to traditional methods for detecting stem cell pluripotency, this method eliminates the destructive and lengthy sample pretreatment required by traditional methods, allowing for the acquisition of biomolecular information from stem cells at different pluripotency states while preserving the spectral signatures within the cells. Therefore, this method enables accurate, non-destructive assessment of the pluripotency state during stem cell differentiation.

[0024] The present invention forms a powerful feature extraction network with CNN and LSTM modules, which can accurately extract local features and time series dependencies in complex spectral data. Moreover, through multi-layer convolution and pooling operations, the model can more comprehensively perceive the key features and dynamic change information in the spectral data, and achieve accurate prediction of the pluripotency state of stem cells. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 A flowchart of an embodiment of the present invention;

[0026] Figure 2 An induction protocol for induced pluripotent stem cells;

[0027] Figure 3 Immunofluorescence images and fluorescence intensity histograms at different time points after replacing RPMI1640 culture medium for induced pluripotent stem cells;

[0028] Figure 4 The single-cell average Raman spectra of six cell states during stem cell culture in RPMI1640 medium;

[0029] Figure 5 Schematic diagram of CNN-LSTM network regression model;

[0030] Figure 6 Loss curve for CNN-LSTM network trained 600 times;

[0031] Figure 7 This is a comparison chart of the predicted and actual values ​​of the iPSC marker SOX2 fluorescence intensity over time;

[0032] Figure 8 This is a comparison chart of the predicted and actual values ​​of the fluorescence intensity of the iPSC marker OCT4 over time. DETAILED DESCRIPTION

[0033] In order to make the purpose and advantages of the present invention more clearly understood, the present invention is described in detail below with reference to the following examples. It should be understood that the following text is only used to describe one or more specific embodiments of the present invention and does not strictly limit the scope of protection of the present invention.

[0034] like Figure 1As shown, the regression prediction method of iPSC pluripotency based on the integration of Raman spectroscopy and CNN-LSTM network regression model includes the following steps:

[0035] Step 1: Acquire a dataset by culturing iPSCs in culture medium and using a Raman spectrometer to collect Raman spectra of the iPSCs during the culture process. The acquisition process is divided into multiple time periods, and spectral data of multiple cells are collected in each time period.

[0036] Step 2: Data preprocessing: removing baseline, removing cosmic rays, reducing noise and normalizing the Raman spectrum data;

[0037] Step 3: Construct and divide the data set. Within a time period, the Raman spectra of the cell samples are divided into six stages in chronological order. The spectral data of each stage is regarded as one class, for a total of six spectral data sets. The processed spectral data are randomly divided into training set, validation set and test set.

[0038] Step 4: Build and train a CNN-LSTM network regression model. The CNN-LSTM network regression model consists of two parts: a CNN and an LSTM. The CNN layer is used to extract the key spectral features of the Raman fingerprint, and the LSTM layer learns the long-term dependencies of the time series, combining the spatial features of the Raman spectrum with the time series dependencies to identify spectra in different time periods.

[0039] Step 5: The trained CNN-LSTM network regression model is used to evaluate the pluripotency of iPSCs. The reserved test set from Step 3, which did not participate in network training, is directly fed into the trained model. This test set is not used in training or hyperparameter tuning, but is used for final model performance evaluation to predict the ability of stem cells to transition between different states. The network evaluates the pluripotency of stem cells, quantifying the pluripotency state as a continuous value, and the regression model predicts the specific value of each state.

[0040] Compared to traditional methods for detecting stem cell pluripotency, the present invention eliminates the destructive and lengthy sample pretreatment processes found in traditional detection methods and can obtain biomolecular information for stem cells in different pluripotency states, preserving the spectral signature information within the cell population. Therefore, the present invention enables accurate, non-destructive assessment of the pluripotency state during stem cell differentiation. Simultaneously, the CNN and LSTM modules form a powerful feature extraction network capable of accurately extracting local features and time series dependencies from complex spectral data. Furthermore, through multi-layer convolution and pooling operations, the model is able to more comprehensively perceive key features and dynamic change information in spectral data, enabling accurate prediction of stem cell pluripotency states. The CNN-LSTM network regression model, through regression prediction of Raman spectra, is able to preserve continuity information between stem cell states, accurately reflect the dynamic transformation process of stem cells, and quantify the gradient of ability to characterize transitions between states, achieving the effect of accurately assessing their pluripotency state and transitional capacity.

[0041] like Figure 2 , in culturing iPSCs, the iPSC culture medium was replaced with RPMI1640 medium at 1, 3, 6, 12, and 24 hours, respectively, thereby affecting the differentiation of iPSCs, such as Figure 3 , iPSC cells were immunofluorescently stained at different time points. iPSCs are characterized by their ability to self-renew indefinitely and generate all cell types of an organism. OCT4 and SOX2 are two essential transcription factors that jointly regulate target genes to maintain the pluripotency of iPSCs. The fluorescence intensity values ​​of the stem cell markers SOX2 and OCT4 at different time points were detected by immunofluorescence staining. Fluorescence intensity can accurately characterize the pluripotency state, so the actual fluorescence intensity was used as the true value for model training. An Olympus fluorescence microscope IX73 was used to capture immunofluorescence images of iPSC cells at six time points; ImageJ was used to calculate the average immunofluorescence intensity of the iPSC colony state. The average fluorescence intensity = the sum of the fluorescence intensity of the colony area / the area of ​​the colony area. The results were plotted as a bar graph to compare the fluorescence intensity at different time points.

[0042] like Figure 4 , based on the single-cell average Raman spectra of six cell states during the culture of stem cells in RPMI1640 medium. When using a Renishaw confocal Raman spectrometer to collect Raman spectral data, the laser wavelength was set to 633nm, the laser power was set to 50%, and the L50× objective lens with a numerical aperture of 0.5 was used to collect 600-1800cm -1Raman spectra were collected over a range of wavenumbers. For each time point, at least 50 cells were sampled for signal acquisition, with at least 15 measurements per cell sample. Data preprocessing was performed using WiRE 4.3 software, including baseline removal, cosmic ray removal, and noise reduction. Before inputting the dataset into the network for training, the spectral data were normalized, scaling the data values ​​to the range [0, 1] to eliminate the potential impact of spectral intensity variations on model results. The processed spectra were divided into six stages (iPSC, 1 hour, 3 hours, 6 hours, 12 hours, and 24 hours) based on the chronological order of cell treatment, constructing a spectral dataset encompassing six cell states. The Raman spectral data were randomly divided into training, validation, and test sets in an 8:1:1 ratio. Stratified sampling was used to ensure a balanced distribution of classes. The resulting training, validation, and test sets can be used for model training, hyperparameter tuning, and performance evaluation, respectively.

[0043] like Figure 5 The CNN-LSTM network regression model consists of a convolutional neural network (CNN) and a long short-term memory network (LSTM). Preprocessed Raman spectral data is used as input, and the fluorescence intensity values ​​of the stem cell markers SOX2 and OCT4 at different time points detected by immunofluorescence staining are used as ground truth input. A CNN is used to extract local features of the spectral data. The input data dimension is the spectrum length. After four layers of convolution with kernel sizes of 12, 14, 16, and 16, the output dimensions are 16, 32, 64, and 128, respectively. Batch normalization (BN) and ReLU activation functions are used after each layer to enhance nonlinear expression capabilities. The data dimension is gradually reduced through a max-pooling layer. The output of the convolutional layer is flattened and passed through a bidirectional LSTM layer. The LSTM layer then performs time series modeling on the 128-dimensional feature sequence, and four layers of bidirectional LSTM capture time series dependencies. The forward LSTM passes information from the beginning to the end of the sequence, capturing past information. The reverse LSTM passes information from the end to the beginning, capturing future information. The number of hidden units is 128, and the output dimension is 256. The outputs of the forward and reverse LSTMs are combined to produce the final output. The LSTM output passes through a dropout layer to prevent overfitting. A fully connected layer (FC) then maps the features to regression values ​​at six time points, corresponding to the pluripotency scores of two markers (SOX2 and OCT4), enabling stem cell pluripotency assessment.

[0044] During the training process, Adam was used as the optimizer to train the model, and the learning rate of Adam was set to 1×10 -5, and the mean square error loss is used as the loss function. The CNN-LSTM network regression model is built through PyTorch, and the divided training set and validation set are put into the network for training. At the end of each round, the model performance is evaluated on the validation set, and the training loss (Train Loss), validation loss (Val Loss), root mean square error (RMSE) and determination coefficient R are recorded. 2 .

[0045] like Figure 6 , the loss of the training and validation data sets gradually decreased, the accuracy steadily increased, and the curves tended to converge, indicating that no overfitting occurred during the training process. Figure 3 From the fluorescence intensity histograms (b) and (c), we observed that the fluorescence intensity of the biomarkers SOX2 and OCT4 in the differentiation process of iPSC decreased with time, indicating that the pluripotency of iPSC gradually decreased during the differentiation process. In order to quantitatively analyze the changes in fluorescence intensity over time, we used the CNN-LSTM network regression model to predict the fluorescence intensity at different time points. The prediction results of the test set were grouped by time point, and the mean of the 6 predicted fluorescence intensity values ​​of the two biomarkers (SOX2 and OCT4) in each category was calculated. Finally, a table with 6 rows and 6 columns was output, where each row corresponds to a time point, the first three columns are the predicted mean values ​​of SOX2, and the last three columns are the predicted mean values ​​of OCT4. Subsequently, the ORIGIN software was used to calculate the average of the three predicted values ​​at each time point and the corresponding actual fluorescence intensity mean, and a line graph was drawn based on these data, as shown in the figure. Figure 7 and Figure 8 In the figure, the gray broken line represents the true fluorescence intensity value at different time points measured by immunofluorescence experiment, while the black broken line represents the fluorescence intensity at different time points predicted by CNN-LSTM network regression model. 2 Indicates the degree of correlation between the predicted value and the true value. A value close to 1 indicates that the predicted value has a high correlation with the true value. The root mean square error RMSE represents the standard deviation of the difference between the predicted value and the true value. The smaller the value, the higher the accuracy of the prediction. During the training process, the R value of the validation set is calculated after each round. 2 and RMSE, used to monitor the performance of the model. During the testing phase, the coefficient of determination R of the test set is calculated. 2and RMSE, used to evaluate the model's final performance. The CNN-LSTM network regression model achieved a coefficient of determination of 0.93256 and a root mean square error of 0.13611 in the quantitative evaluation, indicating that the trends of the predicted and true values ​​were generally consistent. The CNN-LSTM network regression model can output a continuous stem cell pluripotency score, accurately assessing the degree to which iPSCs remain in a specific pluripotent state, capturing the dynamic changes of iPSCs during differentiation, and evaluating their transition states. Combining Raman spectroscopy with the CNN-LSTM network regression model achieved high-precision, continuous assessment of iPSC pluripotency, demonstrating that the CNN-LSTM network regression model overcomes the limitations of traditional classification models, effectively reflects changes in stem cell pluripotency, and provides a powerful tool for stem cell research.

[0046] The foregoing is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art may make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications are also within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained herein shall, unless otherwise specified or limited, be implemented in accordance with conventional means in the art.

Claims

1. A regression prediction method for iPSC pluripotency based on the integration of Raman spectroscopy and a CNN-LSTM network regression model, characterized by: The following steps are involved: Step 1: Acquire a dataset by culturing iPSCs in culture medium and using a Raman spectrometer to collect Raman spectra of iPSCs during the culture process. The acquisition process is divided into multiple time periods, and spectral data of multiple cells are collected in each time period; Step 2: Data preprocessing: removing baseline, removing cosmic rays, reducing noise and normalizing the Raman spectrum data; Step 3: Construct and divide the data set. Within a time period, the Raman spectra of the cell samples are divided into six stages in chronological order. The spectral data of each stage is regarded as one class, for a total of six spectral data sets. The processed spectral data are randomly divided into training set, validation set and test set. Step 4: Build and train a CNN-LSTM network regression model. The CNN-LSTM network regression model consists of two parts: a CNN and an LSTM. The CNN layer is used to extract the key spectral features of the Raman fingerprint, and the LSTM layer learns the long-term dependencies of the time series, combining the spatial features of the Raman spectrum with the time series dependencies to identify spectra in different time periods. Step 5: Use the trained CNN-LSTM network regression model to evaluate iPSC pluripotency. The reserved test set from Step 3, which did not participate in network training, is directly fed into the trained model. This test set is not used in training or hyperparameter tuning and is used for final model performance evaluation. The network evaluates stem cell pluripotency, quantifying the pluripotency state as a continuous value. The regression model then predicts the specific value of each state.

2. The iPSC pluripotency regression prediction method based on the integration of Raman spectroscopy and CNN-LSTM network regression model according to claim 1, characterized in that: In step 1, during the iPSC culture, the iPSC culture medium was replaced with RPMI1640 culture medium at 1, 3, 6, 12, and 24 hours, respectively, and spectra of the iPSC cells at five time points and the original iPSC cells without culture medium replacement were collected at a total of six time points.

3. The iPSC pluripotency regression prediction method based on the integration of Raman spectroscopy and CNN-LSTM network regression model according to claim 2, characterized in that: In step 1, the fluorescence intensity values ​​of stem cell markers SOX2 and OCT4 were detected by immunofluorescence staining; and the average immunofluorescence intensity of the iPSC state was calculated using ImageJ, where average fluorescence intensity = total fluorescence intensity of the colony region / area of ​​the colony region.

4. The iPSC pluripotency regression prediction method based on the integration of Raman spectroscopy and CNN-LSTM network regression model according to claim 3, characterized in that: In step 2, the obtained data is preprocessed, including baseline removal, cosmic ray removal, and noise reduction. Before the data set is input into the network for training, the spectral data is normalized and the data values ​​are scaled to the range of [0, 1].

5. The iPSC pluripotency regression prediction method based on the integration of Raman spectroscopy and CNN-LSTM network regression model according to claim 4, characterized in that: In step 3, the Raman spectrum is divided into six stages according to the processing time sequence, and the Raman spectrum data is randomly extracted into a training set, a validation set, and a test set in a ratio of 8:1:1; the random division process uses stratified sampling, and the divided training set, validation set, and test set are used for model training, hyperparameter tuning, and performance evaluation, respectively.

6. The iPSC pluripotency regression prediction method based on the integration of Raman spectroscopy and CNN-LSTM network regression model according to claim 5, characterized in that: In step 4, the CNN-LSTM network regression model consists of a CNN layer and an LSTM layer; the local features of the spectral data are extracted by CNN, the input data dimension is the spectral length, and after four layers of convolution with convolution kernel sizes of 12, 14, 16 and 16, the output dimensions are 16, 32, 64 and 128 respectively; the LSTM layer performs time series modeling on the 128-dimensional feature sequence, captures the time series dependency through four layers of bidirectional LSTM, and then maps the features to the regression values ​​of 6 time points through a fully connected layer, corresponding to the pluripotency scores of the two markers SOX2 and OCT4, to achieve pluripotency assessment.

7. The iPSC pluripotency regression prediction method based on the integration of Raman spectroscopy and CNN-LSTM network regression model according to claim 6, characterized in that: In step 4, during the training of the CNN-LSTM network regression model, Adam is used as the optimizer to train the model, and the learning rate of Adam is set to 1×10 -5 , and used the mean square error loss as the loss function; built a CNN-LSTM network regression model through PyTorch, and put the training set and validation set after the data set processing in step 3 into the CNN-LSTM network regression model for training. At the end of each round, the model performance was evaluated on the validation set, and the training loss, validation loss, root mean square error and determination coefficient R were recorded. 2 .

Images