Method for scoring adipogenic differentiation of cells based on high throughput and artificial intelligence
Through high-throughput and artificial intelligence technology, deep learning models are used to score cell bright field pictures, solving the problem of time-consuming, labor-intensive and deviation of cell lipid-generating differentiation scores, and achieving fast, batch and accurate scoring effects.
Patent Information
- Application Number
- CN202411486044.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-23
- Publication Date
- 2025-02-11
AI Technical Summary
The prior art has problems such as time-consuming and labor-intensive, observer deviation, and lack of effective methods for cell lipidogenesis differentiation scores in bright field.
Using high-throughput and artificial intelligence methods, the cell bright field pictures are directly scored by quickly taking cell bright field pictures and using deep learning models (combined with yolov7 and ResNet101) for batch learning, supplemented by traditional staining and molecular biology Way.
The rapid, batch and accurate score of cell lipid differentiation is achieved, reducing the difficulty of factor screening and judging harvest time points, and avoiding observer deviation.
Smart Images
Figure CN119418336A_ABST
Abstract
Description
Technical field:
[0001] The present invention belongs to the field of biotechnology, and in particular relates to a method for scoring cell adipogenic differentiation based on high throughput and artificial intelligence. Background technology:
[0002] Cell cultured meat is an emerging technology in the future. It refers to the direct cultivation of livestock and poultry stem cells, which proliferate and differentiate to obtain meat products after tissue engineering processing, so that people can eat pork without raising pigs. In cell cultured meat, the differentiation of cells into muscle and fat is a key link. Among them, fat is an important nutrient and flavor component that is indispensable to meat. In this process, the evaluation of cell adipogenesis is traditionally done by staining, such as oil red staining, fluorescent staining (Bodipy, Nile red, etc.) to obtain lipid staining images of cell differentiation, and analyze the images (counting lipid droplets or staining area) to determine the cell differentiation effect, or by molecular biology experiments, extracting key genes and proteins in the cell adipogenesis process, and analyzing their changes to obtain cell differentiation effects. Both methods are time-consuming and labor-intensive in screening factors that induce cell differentiation effects, and the former often produces deviations due to different observers.
[0003] Artificial intelligence has been a very popular research direction in recent years. It has been widely used in computer vision. Deep learning, with its efficient image feature recognition and prediction capabilities, has provided new opportunities and solutions for the study of cell biology. It can replace the above existing methods to directly evaluate the adipogenic differentiation of cells in images. Studies have shown that deep learning can be used to directly count lipid droplets in cells for evaluation. However, this application is mostly used for 2D cells, and there are few applications in the adipogenic differentiation of 3D suspended cells.
[0004] In response to the above problems, the efficiency of developing methods for scoring cell adipogenic differentiation based on high throughput and artificial intelligence is improved. The main technical barriers in practice are as follows:
[0005] 1) The sample size needs to be sufficient; the cell data in these samples need to be presented in the form of single cells without clumps;
[0006] 2) Model selection and loss function determination: Model identification is to lay the foundation for predicting values. In this part, the currently mature network architecture should be used. In the prediction part, the input cells of the model range from D0 to D4, and the value range is relatively wide. The loss function should be determined to be consistent with the trend of the model training process.
[0007] 3) How the model is used in practice: The performance of the model can be evaluated by indicators, but due to the wide range of values and the diversity of biological data, indicators cannot directly explain their effects. Given sufficient data, how can the model determine the evaluation of samples in practice?
[0008] 4) Effect of different batches of cells on adipogenic differentiation scores during production: There are few reports on models that directly evaluate adipogenic differentiation using bright field images of cells. Summary of the invention:
[0009] In view of the above defects or improvement needs of the prior art, the present invention provides a method for scoring cell adipogenic differentiation based on high throughput and artificial intelligence, which aims to directly score cell bright field images by using high throughput rapid shooting and artificial intelligence batch learning of images, while traditional methods, staining and molecular biology methods will be used as auxiliary means. It can be used to quickly screen factors that promote differentiation of cells and help us quickly determine the time point for harvesting cell adipogenic differentiation.
[0010] The present invention provides a method for scoring cell adipogenic differentiation based on high throughput and artificial intelligence, which solves the current problem of lack of cell adipogenic differentiation scoring under bright field, has the advantages of being fast, batch and accurate, and provides guidance for product harvesting (determining whether the cells meet the harvesting requirements) and factor screening.
[0011] To achieve the above objectives, this application proposes the following technical solutions:
[0012] 1) Dataset acquisition:
[0013] Bright field images of cells from D0 to D4 were acquired. Under the high-throughput scheme, the images were taken with the control group as the unified standard (here D0). The cell fluorescence intensity values were also acquired with the control group D0 as the standard. The fluorescence intensity values corresponded to each single cell and were named. The sizes of all images were kept consistent during model training, and data enhancement was used to expand the data set.
[0014] 2) Network framework construction:
[0015] The yolov7 network architecture for target detection (recognition model) is an open source structure that can be directly obtained (https: / / github.com / bubbliiiing / yolov7-pytorch). The network structure for fluorescence intensity value prediction (prediction model) uses ResNet101 as the basic architecture and is enhanced by introducing the Squeeze-and-Excitation (SE) module. The SE module enhances the learning of important features by adaptively adjusting the weights of feature channels. It is set in front of the fully connected layer of resnet101, and then the fully connected layer is modified, and a Dropout regularization layer is added to prevent overfitting. At the same time, the number of output channels is changed to 1 to complete the numerical prediction.
[0016] 3) Loss function, optimizer and learning rate setting:
[0017] In the target detection part, the yolov7 loss function is introduced as its own YOLOLoss loss function. In the part of predicting the fluorescence intensity value, according to the HuberLoss formula, the Huber loss function with an adaptive parameter δ is adopted to improve the robustness of the model to outliers and reduce the calculation of partial derivatives. The AdamW optimizer is used to adjust the model weights, with an initial learning rate of 0.0002 and a weight decay of 0.001. The ReduceLROnPlateau strategy is used to adaptively adjust the learning rate. When the verification loss has not been significantly improved for a long time, the learning rate is reduced by 10 times.
[0018] 4) Model evaluation:
[0019] The final effect of the model is to identify a complete cell bright field image obtained by high-throughput photography, count the number of cells in the image, predict the fluorescence intensity value of each cell, and calculate the average fluorescence intensity value and standard deviation of the image to evaluate the effect of the model in practice. The three indicators MSE, RMSE, R 2 A preliminary evaluation is made on the corresponding numerical results.
[0020] Furthermore, different sample inoculation gradients were set for the 96-well plate of the shooting system. Each well plate was inoculated according to the inoculation number of 1, 6, 12, 24, and 48. Three groups of samples with different differentiation states were set (3 parallel samples in each group) and inoculated separately. Only one sample was inoculated on each plate. The error was compared between the average fluorescence intensity value calculated according to the model's prediction and the actual fluorescence intensity value. The actual performance of the model was determined by the difference in the error size and the actual workload.
[0021] Furthermore, whether the model has the ability to distinguish different samples needs to be evaluated. Three groups of samples with different differentiation states are set up, and they are inoculated on the same plate according to the number of inoculation wells obtained in the previous experiment. The error between the model prediction value and the true value is also compared. Based on the two verification experiments, it is determined how the model should be used in practice.
[0022] As the first method for scoring cell adipogenic differentiation, the mean fluorescence intensity value was converted into a percentage score of 0-100:
[0023]
[0024] x i represents the average fluorescence intensity of the i-th image, x max and x min Respectively represent the maximum and minimum values in the range of recorded average fluorescence intensity; y i Represents the converted value on a percentage scale.
[0025] As a second method for scoring cell adipogenic differentiation, the mean fluorescence intensity was divided into levels using the following formula:
[0026] y=0.5226x-52.014
[0027] Among them, x represents the average fluorescence intensity of the image, y represents the value after division, and the gear position is determined according to the value of y, where:
[0028] y∈[100,+∞), represents gear A; y∈[67,100), represents gear B; y∈[51,67), represents gear C; y∈[13,51), represents gear D; y∈[0,13), represents gear E; y∈(-∞,0), represents gear F.
[0029] Beneficial effects:
[0030] Compared with the previous quantitative analysis of cell adipogenic differentiation, whether it is oil red, fluorescent staining image analysis or molecular biology (QPCR\Wb) gene or protein level analysis, it is time-consuming and labor-intensive, and staining image analysis is prone to observer bias. The present invention solves the above problems and has the advantages of being intuitive, accurate, and batch-based, making it more convenient to screen factors that induce differentiation in the future. Description of the drawings:
[0031] For ease of explanation, the present invention is described in detail with reference to the following specific implementations and accompanying drawings.
[0032] Figure 1 It is a flowchart of the present invention
[0033] Figure 2 This is the original data display diagram of the present invention
[0034] Figure 3 It is the network architecture model of the model prediction part of the present invention
[0035] Figure 4 This is the actual prediction effect diagram of the present invention
[0036] Figure 5 This is a schematic diagram of the error results of the model sample of the present invention inoculated with 12 holes
[0037] Figure 6 This is the final rating chart of the model of the present invention
[0038] Figure 7 This is the correlation result diagram of the average fluorescence intensity value and the differentiation days of the present invention Specific implementation method:
[0039] The present invention is described below in conjunction with the accompanying drawings and embodiments:
[0040] The present invention mainly combines the target detection model with the prediction numerical regression model to propose a cell bright field adipogenic differentiation evaluation method. In the present invention, the target detection model adopts the yolov7 model, and the prediction regression adopts the resnet101 model. The two are combined to obtain a model for directly evaluating adipogenic differentiation of bright field cell images obtained under high-throughput shooting.
[0041] Combination Figure 1 According to the model description, the specific operation is to establish a high-throughput solution, shoot D0-D4 cells with unified shooting and analysis parameters, collect bright field images as data sets, and obtain the fluorescence intensity value of each cell under the corresponding bright field when shooting in the fluorescence channel; the cell bright field image is divided into two parts, the complete bright field image is used as the data set of the recognition model part, and the bright field part of a single cell is used as the data set of the prediction model part. The models of the two parts are trained separately to obtain different model weights. Each picture of cell recognition is annotated by EasyDL and the label 'cell' is set. Flip_Only_BBoxes and Solarize_add are used for data enhancement, and three methods are traversed in parallel to enter the yolov7 network training; in numerical prediction, cells are input as single-cell bright field images, and the name of each image is named after the fluorescence intensity value of the corresponding cell, and it is used as the label (label) as follows Figure 2 The raw data shown shows that the training process of the resnet101 model is Figure 3 illustrate:
[0042] Based on the network construction steps of pytorch, define the CellDataset dataset and set its attribute transform. The image size used is uniformly set to transform.Resize. At the same time, add data enhancement methods, including
[0043] RandomHorizontalFlip / RandomVerticalFlip, RandomRotation with a maximum angle of 20°, RandomAffine, and 20% ColorJitter. The above operations are all part of the dataset definition process. The model uses the pytorch tool library nn.Module to define the prediction model ResNet101 network structure, using the network's pretrained mode. The weights are trained on the ImageNet dataset. The input data meets the model adaptation requirements. The pixels of its three channels are standardized to conform to the ImageNet dataset. Transform.Normalize determines the mean value mean = [0.485, 0.456, 0.406] and the standard deviation std = [0.229, 0.224, 0.225]:
[0044]
[0045] mean R =0.485
[0046] mean G =0.456
[0047] mean B =0.406
[0048] std R =0.229
[0049] std G =0.224
[0050] std B =0.225
[0051] R / G / B, r / g / b, r' / g' / b' all correspond to the three-channel RGB images of the image. The input image is aligned with the mean and standard deviation of IamgeNet.
[0052] The loss function of the model is HuberLoss, which is defined by the formula, and the key hyperparameter δ is set to 0.5.
[0053] The model predicts the cell fluorescence intensity value based on the ResNet101 algorithm and network structure layer. In addition to the basic con2v, relu, and maxpooling, the network structure includes multiple residual blocks combined through Bottleneck, plus GAP, and finally connected to a Fully Connected
[0054] Layer (FCL) completes the output. The basic structure of the network is:
[0055]
[0056] The middle Conv2, Conv3, Conv4, and Conv5 are multiple residual units combined in the Resnet network structure model. Each module will complete a downsampling to reduce the feature by half. The first downsampling is done in Conv1, and the downsampling is completed by adjusting the stride to 2. This network structure has been trained on the large dataset ImageNet, and its pre-trained weights can be directly obtained. The final output layer of Resnet101 is a fully connected layer with a Softmax probability output function. Its original purpose is to output the probability results between 0 and 1 for different categories in multi-classification tasks. Here, the number of fully connected output channels is changed to 1, the Softmax layer is removed, and the SE attention mechanism and Dropout layer are added. The SE attention mechanism is added before the fully connected layer. The detailed process changes are as follows:
[0057] The input single-cell bright field image is resized to a uniform size to obtain X, and then the image is obtained through data enhancement and tensorization.
[0058] X 3×128×128 ;
[0059] Normalization: X 3×128×128 →Tc=(Tc-mean c ) / std c c:channels (RGB three channels);
[0060] After Tc passes through resnet101, its feature map x is extracted. The addition of the SE layer attention mechanism is to adaptively adjust the weight of each channel, thereby enhancing important features and suppressing unimportant features. In the following calculation, x(gap) is a one-dimensional vector obtained by taking the average value of the global average pooling layer (GAP) on each channel, and its shape is b×2048, b represents the batch size, and 2048 is the number of channels. W1 is the first weight matrix of the SE layer, and its shape is 128×2048. Here, the input 2048-dimensional feature vector is reduced to 128, and W2 is the second weight matrix of the SE layer, and its shape is 2048×128. It acts in the opposite way and is dimensionalized to 2048. Y is the channel weight vector, which is calculated by the fully connected layer and activation function of the SE layer. The final Xscale is the feature map after the channel weight is recalibrated.
[0061] SE layer:x(gap)=GAP(x)→R b×2048
[0062] Y=Sigmoid(W2*ReLu(W1*x(gap))
[0063] W1=R128×2048, W2=R2048×128
[0064] X scale =x*y
[0065] A Dropout layer is added to the final output of Resnet101, and the dropout rate parameter is set to 0.5. Half of the neurons will be "turned off" in each forward pass to prevent overfitting, and a fully connected layer is defined to recalibrate the feature map Xscale to 256. The Dropout layer is used again, and the same dropout rate is set to 0.5 to reduce the dimension of its output channel to 1, completing the predicted output value Z of the final model; Xdrop1 is the first layer feature vector after Dropout (0.5); A1 and A2 are the weight matrices acting on the first and second fully connected layers respectively; b1 and b2 are the biases of the first and second fully connected layers respectively; H1 is the activated first fully connected layer (Relu activation); Hdrop2 is the feature vector of the second Dropout;
[0066] Xdrop1=Dropout(Xscale,p=0.5)
[0067] H1=Relu(A1*Xdrop1+b1)
[0068] Hdrop2=Dropout(H1,p=0.5)
[0069] Z=A2*Hdrop2+b2
[0070] After the above process, the model completes the prediction of the fluorescence intensity value of a single cell.
[0071] The two parts of the model are integrated, that is, the weights obtained from the two parts of training. The model can realize the cell recognition and fluorescence intensity value prediction under the whole bright field. By adding the calculation of mean and std, the average fluorescence intensity value and standard deviation are obtained. The average fluorescence intensity value represents the quality of cell adipogenesis, while the standard deviation represents the differentiation difference between cells in the picture. The actual effect is as follows Figure 4 As shown, the original image is input to obtain the predicted image of the model.
[0072] After the two models were integrated, the overall model was evaluated. Different samples (differentiation states) were inoculated in different well plates, and the difference between the mean fluorescence intensity values predicted by the model and the actual values was compared. Figure 5 The results showed that 12 inoculations were sufficient for the model to evaluate a sample quickly, accurately, and stably. At the same time, the model could correctly distinguish different samples by inoculating samples of different differentiation states in 12 wells on the same plate. This confirmed that our model needs to inoculate 12 wells to evaluate each sample in practical applications.
[0073] Finally, according to Figure 6 The displayed scoring and rating table can score and rate the sample cells by converting their average fluorescence intensity values. Based on the actual situation, if the score exceeds the limit, the adipogenic differentiation effect of the cells is judged to be excellent, and the product can be collected and processed.
[0074] The hardware system used in this embodiment is Windows 10, the processor is Intel(R) Xeon(R) Silver4210R CPU, the maximum main frequency is 2.40GHz, the number of cores is 20, the above software compilation environment is Pycharm, the programming language used is python3.7, and the Pytorch deep learning framework is adopted.
[0075] Attached is a Chinese-English comparison table of proper nouns:
[0076] ReduceLROnPlateau: adaptive learning rate adjustment
[0077] RandomHorizontalFlip: Random horizontal flip
[0078] RandomVerticalFlip: Random vertical flip
[0079] RandomRotation: Random rotation within a certain angle range
[0080] RandomAffine: Random affine transformation
[0081] ColorJitter: Color jitter
[0082] Theoretical basis: Combination Figure 7 ,There was a significant correlation between the average fluorescence intensity value and the differentiation days during the differentiation process, [D0, D1, D2, D3, D4] = [99.46, 124.81, 196.63, 228.40, 290.67], and the straight line fitting formula was: Y = 48.60x + 90.79, where x represented the differentiation days, Y represented the average fluorescence intensity value, the determination coefficient R2 = 0.98, and the Pearson coefficient = 0.99.
[0083] It should be noted that the above-described embodiments only express the main implementation methods of the present invention, and the description thereof cannot be understood as limiting the scope of the present invention. It should be pointed out that the recognition and prediction of the model can be completed by other similar model network structures, and for ordinary technicians in the field of cell biology, several improvements can be made without departing from the concept of the present invention, which should all fall within the scope of protection of the present invention.
Claims
1. A method for scoring cell adipogenic differentiation based on high throughput and artificial intelligence, characterized in that: The steps include: S1. Construct a mean fluorescence intensity prediction model, specifically including: (1) Obtaining a data set, and obtaining bright field images of cells at different stages through high-throughput photography; and obtaining the fluorescence intensity value of each cell in the bright field image at different stages and the average fluorescence intensity value of all cells in the bright field image based on analysis; (2) Model establishment: the first part is the recognition model: identifying, counting, and segmenting each cell under bright field; the second part is the prediction model: predicting the fluorescence intensity value of cells under bright field; (3) Model training: for the recognition model, the cell recognition, counting and segmentation parts are trained separately; for the prediction model, the cell fluorescence intensity value is trained. The two parts are trained separately and then integrated to obtain a complete model for segmenting single cells in bright field images and predicting the fluorescence intensity value of each recognized cell. (4) Output the average fluorescence intensity value of all cells in the bright field image; S2, high-throughput acquisition of bright field images of cells to be predicted, input into the mean fluorescence intensity prediction model, and obtain the mean fluorescence intensity value; S3. Divide the levels based on the mean fluorescence intensity value to obtain the cell adipogenic differentiation score.
2. The method according to claim 1, characterized in that: In the acquisition of the data set, high-throughput shooting: using undifferentiated cells D0 as a standard, unifying the shooting parameters of cells under bright field, Bodipy fat staining, and Hoechst cell nucleus staining from D0 to D4, and using D0 as a standard, unifying the parameters of cells under Bodipy fluorescence shooting and analysis from D0 to D4; obtaining bright field images of cells under different differentiation days, and obtaining the fluorescence intensity value of each cell in the bright field images under different differentiation days according to the analysis, as well as the corresponding average fluorescence intensity value; D1, D2, D3, and D4 represent the first day of differentiation, the second day of differentiation, the third day of differentiation, and the fourth day of differentiation, respectively.
3. The method according to claim 2, characterized in that: Cells in different differentiation states correspond to cells collected at different differentiation days. During the cell differentiation process, feeding is performed every other day, and the overall differentiation cycle is 4 days.
4. The method according to claim 1, characterized in that: The recognition model adopts the yolov7 algorithm and network architecture. The complete bright field image is used as the data set of the recognition model. It is marked through the EasyDL platform, and data enhancement is used. Parallel enhancement and flipping are selected to expand the data volume. The training: validation: test = 8:1:1 is set, and the SGD optimizer is used. The weight decay parameter is 5e-4, the learning rate is 1e-2, and the cosine annealing learning rate scheduling cos method is used.
5. The method according to claim 1, characterized in that: The prediction model is a regression task, where the single-cell bright field image is the input, and the fluorescence intensity value corresponding to the cell is the feature label. All input images are processed into a uniform resolution size. Data enhancement is used to add random horizontal and vertical flips, random rotations within 20 degrees, affine transformations, and color transformations to expand the data volume, setting the training:test ratio to 8:
2.
6. The method according to claim 5, characterized in that: The prediction model uses the pytorch tool library nn.Module to define the prediction model ResNet101 network structure and uses the network's pretrained mode. The weight is trained on the ImageNet dataset. The input data meets the model's adaptation requirements. The pixels of its three channels are standardized to conform to the ImageNet dataset. The final output layer of Resnet101 is a fully connected layer. The SE attention mechanism and Dropout layer are added. The SE attention mechanism is added before the fully connected layer to adaptively adjust the weight of each channel. The Dropout layer sets the dropout rate parameter to 0.5, and half of the neurons will be "turned off" in each forward pass to prevent overfitting.
7. The method according to claim 5, characterized in that: The choice of prediction model loss function is determined as HuberLoss, the formula of HuberLoss is: δ represents a hyperparameter, y represents the target value, and f(x) represents the predicted value of the model.
8. The method according to claim 5, characterized in that: The model is evaluated by the mean square error (MSE), root mean square error (RMSE), and coefficient of determination (R2), among which: y i is the actual value of the observed ith dependent variable, is the predicted value corresponding to the observed dependent variable, is the average of the actual values of the dependent variable, and n is the number of samples.
9. The method according to claim 1, characterized in that: Different sample inoculation gradients were set for the 96-well plate of the shooting system. Each well plate was inoculated with 1, 6, 12, 24, and 48 inoculation holes. Three groups of samples with different differentiation states were set and inoculated separately. Only one sample was inoculated on each plate. The error was compared between the average fluorescence intensity value calculated according to the model's prediction and the actual fluorescence intensity value. The actual performance of the model was determined by the difference in the error size and the actual workload.
10. The method according to claim 1, characterized in that In step S3: convert the mean fluorescence intensity value into a percentage score of 0-100 as the cell adipogenic differentiation score: x i represents the average fluorescence intensity of the i-th image, x max and x min Respectively represent the maximum and minimum values in the range of recorded average fluorescence intensity; y i Express the converted value on a percentage scale; or, The mean fluorescence intensity was calculated by the following formula and divided into different levels as the score of cell adipogenic differentiation: y=0.5226x-52.014 Among them, x represents the average fluorescence intensity of the image, y represents the value after division, and the gear is determined according to the value of y, where: y∈[100,+∞), represents gear A; y∈[67,100), represents gear B; y∈[51,67), represents gear C; y∈[13,51), represents gear D; y∈[0,13), represents gear E; y∈(-∞,0), represents gear F.