A prediction method and device for screening transfected cells, an electronic device, and a storage medium

Abstract

The present invention provides a prediction method and device for cell screening after transfection, an electronic device, and a storage medium, which relate to the field of computer technology; it includes: obtaining data of high-quality cell screening projects as an original data set and preprocessing it to verify the rationality of the data preprocessing strategy, so as to obtain feature data; constructing a fusion sequence self-attention network model, and inputting the obtained feature data into the fusion sequence self-attention network model to obtain a prediction result. The present invention uses data of high-quality cell screening projects as the original data set and adopts SMOTE to solve the problem of data label imbalance in the original data set; through the FSAN model, combining the long-term dependence processing ability of LSTM and the local pattern capturing ability of CNN, it provides rich feature representations for the Encoder module; through the self-attention layer, it identifies and focuses on the complex relationships between the data features extracted by LSTM and CNN.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and particularly to a prediction method and device for screening transfected cells, an electronic device, and a storage medium. Background Art

[0002] In the modern biomedical field, the development of biopharmaceutical technology has greatly promoted the progress of the treatment of various diseases. As an important means of biopharmaceuticals, cell engineering pharmaceuticals have formed a very large industrial scale. Cell engineering pharmaceuticals is a method of using genetic engineering or cell culture technology to transform cells so that they can produce drugs. Therefore, obtaining high-quality transformed cells is crucial, which will directly determine the quality and efficiency of drug production. The currently most commonly used cell transformation method is the transfection method, in which specific exogenous nucleic acids are introduced into host cells so that the cells can produce specific drugs. However, due to the randomness of the position of exogenous nucleic acids inserted into the gene sequence of host cells, after transfection, it is necessary to screen out high-quality cells with good growth conditions, high expression levels, and stable expression levels from a large number of cells.

[0003] The currently widely used screening methods mainly include two stages. The first stage is the deep well plate stage. A large number of cells obtained by transfection are grouped and then placed in deep well plates for culture. During the culture period, cells with poor growth conditions are screened out, and finally only cells with good performance are retained. The second stage is the shake flask stage. The cells screened in the deep well plate stage are placed in shake flasks for continued culture. During this period, culture data such as cell concentration and average cell particle size are measured at regular intervals, and cells that do not meet the conditions are continuously screened out according to the measured data. Finally, only a few groups of cells with excellent performance are retained for the subsequent process.

[0004] The above screening methods rely on time-consuming culture, which has become a bottleneck restricting the efficiency of cell engineering pharmaceuticals. Long-term culture and screening require a large number of professional equipment and trained operators, resulting in high costs. The frequent intervention of manual operations also greatly increases the risk of cell contamination and experimental errors. The cell screening is determined by operators, and the screening results are easily affected by the subjective judgment of operators, which greatly weakens the reliability and consistency of the screening results. Summary of the Invention

[0005] The purpose of the present invention is to develop an automated and efficient cell screening prediction method and device, an electronic device, and a storage medium for the problems of excessive screening process time, reduction of labor costs and manual subjective deviation in the screening process, and improvement of screening quality. The specific technical solutions are as follows.

[0006] The present invention provides a prediction method for screening transfected cells, including the following steps:

[0007] Step 1: Obtain the data of high-quality cell screening projects as the original dataset;

[0008] Step 2: Preprocess the data in the original dataset, and use correlation analysis and significance testing to verify the rationality of the data preprocessing strategy to obtain feature data;

[0009] Step 3: Construct a fusion sequence self-attention network model, input the obtained feature data into the fusion sequence self-attention network model, and obtain the prediction result.

[0010] Further, the original dataset includes multiple pieces of data, and each piece of data includes a time series with a length of 10 days and six key feature indicators; the six key feature indicators are viable cell concentration, total cell concentration, total cell number, average particle size, average roundness, and aggregation rate.

[0011] Further, the acquisition process of a single piece of data is as follows:

[0012] ①. In the deep-well plate stage, data is collected only once after four days of culture to obtain one piece of data;

[0013] ②. In the shake flask stage, one piece of data is collected on the first, third, fifth, seventh, eighth, ninth, tenth, twelfth, and fourteenth days of the shake flask stage to obtain nine pieces of data; cells that do not meet the requirements are screened out according to the growth conditions of the cells, and the incomplete time series is filled with 0;

[0014] ③. Concatenate the one piece of data obtained in the deep-well plate stage with the nine pieces of data obtained in the shake flask stage to obtain a time series with a length of 10 days.

[0015] Further, the specific process of preprocessing the data in the original dataset is as follows:

[0016] Select any two pieces of data in the original dataset with a label of 1, and use the median interpolation method to standardize the two pieces of data selected in the original dataset with a label of 1 to generate new sample data marked as ; Standardize all the data in the original dataset to obtain the standardized original dataset; ; Add a random perturbation of ±0.2 to the standardized data, and generate four new sample data from each new sample data

[0017] to obtain the preprocessed dataset;

[0018]

[0018] Perform Pearson correlation analysis on the standardized original dataset and the preprocessed dataset and Testing is performed to obtain feature data, which includes the correlation changes between different features and the values of its corresponding features.

[0019] Furthermore, the data in the original dataset is stored in a CSV file;

[0020] The results obtained by using Pearson correlation analysis and testing are all stored in a CSV file.

[0021] Furthermore, the process of constructing the fusion sequence self-attention network model is as follows:

[0022] Configure the long short-term memory neural network model as a two-layer structure, with each layer having a variable configuration of 16 to 64 hidden units, and the dropout function is set to 0.5;

[0023] Configure the convolutional neural network as a variable configuration containing 1 to 2 convolutional layers, use a convolutional kernel of size 2, and do not use a pooling layer to capture the dynamic changes of cells in a short time;

[0024] Based on the Encoder module of the self-attention mechanism, use single-head attention, followed by a single-layer feed-forward neural network; among them, the Encoder module uses the single-head attention mechanism to process the input information; the implementation method of single-head attention is to obtain query, key, and value matrices by linearly transforming the input; the attention mechanism calculates the similarity between the query and the key, assigns different weights to the inputs at different positions, so as to capture the correlation and dependence between the inputs at different positions; after the attention output, a single-layer feed-forward neural network is used to perform further non-linear transformation on the data to enrich the feature expression;

[0025] Concatenate the output of the last layer of the LSTM and the flattened output of the CNN as the input of the Encoder;

[0026] The output of the Encoder passes through a Softmax activation function to generate the final binary classification prediction fusion sequence self-attention network model.

[0027] Furthermore, the specific process of inputting the obtained feature data into the fusion sequence self-attention network model for prediction is as follows:

[0028] The preprocessed dataset is divided into a training set, a validation set, and a test set according to a ratio of 6:2:2, and all random seeds are set to 202;

[0029] The training set is trained using the Encoder module of the fusion sequence self-attention network model. When training the training set using the Encoder module of the fusion sequence self-attention network model, the Adam optimizer and the cross-entropy loss function are used, and the accuracy, precision, recall, and F1 score are used as evaluation metrics for the fusion sequence self-attention network model to obtain the prediction results.

[0030] The present invention also provides a post-transfection cell screening prediction device for implementing the prediction method of post-transfection cell screening as described above, including:

[0031] An original data set import unit for obtaining the original data set of cell screening;

[0032] An original data set preprocessing unit for preprocessing the original data set to obtain feature data;

[0033] A cell screening prediction unit for inputting the preprocessed feature data into the fusion prediction model for prediction, obtaining the prediction results, and exporting the prediction result file.

[0034] The present invention also provides an electronic device, including a processor and a memory connected to each other;

[0035] The memory is used to store a computer program that supports the electronic device to execute the prediction method of the cell screening as described above, and the computer program includes program instructions;

[0036] The processor is configured to call the program instructions to execute the prediction method of the cell screening as described above.

[0037] The present invention also provides a storage medium, the storage medium stores a computer program, the computer program includes program instructions, and the program instructions cause the processor to execute the prediction method of the cell screening as described above when being executed by the processor.

[0038] Applying the technical solution of the present invention has at least the following beneficial effects:

[0039] (1) The present invention pre - processes the original data set by using the data of high - quality cell screening projects as the original data set, and adopts the Synthetic Minority Over - sampling Technique (SMOTE) to solve the problem of data label imbalance in the original data set, and then alleviates the problem of small data volume through data augmentation. Then, through the FSAN model, combining the long - term dependence processing ability of LSTM and the local pattern capturing ability of CNN, rich feature representations are provided for the Encoder module. Through the self - attention layer, the model can identify and focus on the complex relationships between the data features extracted by LSTM and CNN. In order to ensure the effectiveness and practicality of the model, considering the model complexity and data characteristics, special attention is paid to the design of the fused sequence self - attention network model and the selection of parameters during construction. The constructed fused sequence self - attention network model can effectively process data with different time - series lengths, providing a trade - off solution for researchers between prediction accuracy and shortening the cell screening time. This method highlights the potential of effectively using deep learning techniques for time - series analysis in cell screening.

[0040] (2) The present invention proposes a cell screening prediction device capable of implementing the cell screening prediction method.

[0041] (3) The present invention proposes an electronic device capable of storing and executing a computer program of the cell screening prediction method.

[0042] (4) The present invention proposes a storage medium capable of storing a computer program containing specific program instructions, and when the program instructions are executed by a processor, the processor executes the cell screening prediction method.

[0043] In addition to the purposes, features and advantages described above, the present invention has other purposes, features and advantages. The following will refer to the drawings for a further detailed description of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The drawings forming a part of this application are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0045] Figure 1 is a flowchart of a cell screening prediction method according to Embodiment 1 of the present invention;

[0046] Figure 2 is a structural diagram of the FSAN fused sequence self - attention network model in Embodiment 1 of the present invention;

[0047] Figure 3Schematic diagram of the prediction process of the cell culture prediction device according to Embodiment 2 of the present invention. Detailed implementation manners

[0048] To make the above objects, features, and advantages of the present invention more clearly understandable, the following will describe the detailed implementation manners of the present invention with reference to the accompanying drawings. It should be noted that the accompanying drawings of the present invention are all in simplified forms and use non-precise scales, only for facilitating and clearly assisting in the description of the embodiments of the present invention; the several mentioned in the present invention are not limited to the specific quantities in the accompanying drawing examples; the orientation or positional relationships indicated by 'front','middle', 'back', 'left', 'right', 'up', 'down', 'top', 'bottom','middle', etc. in the present invention are all based on the orientation or positional relationships shown in the accompanying drawings of the present invention, and do not indicate or imply that the devices or components referred to must have a specific orientation, nor can they be understood as a limitation to the present invention. Embodiment 1:

[0049] A prediction method for screening transfected cells provided by the present invention, see Figure 1 shown, including the following steps:

[0050] Step 1: Obtain data of high-quality cell screening projects as the original data set.

[0051] Preferably, the original data set includes data of 9 high-quality cell screening projects that have been put into drug production.

[0052] Furthermore, the original data set includes 217 pieces of data of various product types such as monoclonal antibodies, recombinant proteins, and vaccines. Each piece of data includes a time series with a length of 10 days and six key feature indicators; the six key feature indicators are viable cell concentration (F1), total cell concentration (F2), total cell number (F3), average particle size (F4), average roundness (F5), and aggregation rate (F6).

[0053] Even further, the data in the original data set is stored in a CSV file. Specifically, with viable cell concentration (F1), total cell concentration (F2), total cell number (F3), average particle size (F4), average roundness (F5), and aggregation rate (F6) as the headers, an English comma is used as the separator between the viable cell concentration (F1), total cell concentration (F2), total cell number (F3), average particle size (F4), average roundness (F5), and aggregation rate (F6) in each row of data.

[0054] Preferably, the acquisition process of a single piece of data is as follows:

[0055] ①. In the deep well plate stage, since the amount of cell culture medium is small, sampling will cause relatively large losses. Therefore, data is collected only once after 4 days of culture to obtain one piece of data;

[0056] ②. In the shake flask stage, one data is collected for each of the first, third, fifth, seventh, eighth, ninth, tenth, twelfth, and fourteenth days of the shake flask stage, obtaining nine data; and cells that do not meet the requirements are screened out according to the growth of the cells. Therefore, some cells do not have complete culture data, and the incomplete time series is filled with 0s for this

[0057] ③. Concatenate the one data obtained in the deep well plate stage with the nine data obtained in the shake flask stage to obtain a time series with a length of 10 days.

[0058] Step 2. Preprocess the data in the original dataset and use correlation analysis and significance testing to verify the rationality of the data preprocessing strategy to obtain feature data.

[0059] Preferably, the specific process of preprocessing the data in the original dataset is as follows:

[0060] S2.1. Use the SMOTE algorithm to balance the labels of the data in the original dataset to obtain the standardized original dataset; the specific method is:

[0061] Select any two data with a label of 1, and use the median interpolation method to standardize the two selected data with a label of 1 to generate a new sample data labeled ;

[0062] Perform standardization processing on all the data in the original dataset to obtain the standardized original dataset;

[0063] The expression of the SMOTE algorithm is as follows:

[0064] ;

[0065] where is the generated new sample data; is a sample in the minority class, and the minority class represents a class without complete culture data; is one of the k-nearest neighbor model samples of is a random number, taking values in the interval, used to control the random offset of the new sample.

[0066] S2.2. Perform data augmentation processing on the new sample data ; the specific method is:

[0067] Add a random perturbation of ±0.2 to the standardized data, and for each new sample data Generate four new sample data , and obtain the preprocessed dataset; by expanding the original dataset, the adaptability and generalization ability of the FSAN fusion sequence self-attention network model to the new sample data are improved.

[0068] S2.3. Perform Pearson correlation analysis and test (unequal variances) on the standardized original dataset and the preprocessed dataset to obtain the correlation changes between different features and the values of their corresponding features to judge the rationality of preprocessing the data in the original dataset.

[0069] The correlation changes between different features obtained by Pearson correlation analysis are expressed as follows:

[0070] ;

[0071] where is any one of the key feature indicators, is any one of the key feature indicators, , is the mean value, is the mean value, is the total number of times of correlation analysis of six key feature indicators, is the number of key feature indicators (i.e., is 6);

[0072] The expression of the test (unequal variances) is as follows:

[0073] ;

[0074] where and are the mean values of the samples in the original dataset and the preprocessed dataset respectively, and are the variances of the samples in the original dataset and the preprocessed dataset respectively, and are the sample sizes of the samples in the original dataset and the preprocessed dataset respectively.

[0075] The results obtained by Pearson correlation analysis and test (unequal variances) are all stored in a CSV file. Specifically, using statistic, The value is the table header, and in each row of data statistics, between values are separated by English commas. Among them, the value represents the probability of observing the statistic or a more extreme case when the null hypothesis (i.e., there is no correlation between two variables) is true. The value and the statistic are used together to evaluate whether the correlation between different features is statistically significant.

[0076] The correlation change between different features and the statistic, the corresponding relationship of the value is shown in Table 1.

[0077] Table 1: The corresponding relationship between the correlation change between different features and the statistic, the value

[0078] .

[0079] Step 3: Construct an FSAN fusion sequence self-attention network model, input the obtained feature data into the FSAN fusion sequence self-attention network model, and obtain the prediction result.

[0080] Preferably, as shown in Figure 2 , the process of constructing the FSAN fusion sequence self-attention network model is as follows:

[0081] Step 3.1: Configure the long short-term memory neural network model (i.e., LSTM) as a two-layer structure, with a variable configuration of 16 to 64 hidden units in each layer, and the dropout function is set to 0.5 to avoid overfitting;

[0082] Step 3.2: Set the convolutional neural network model (i.e., CNN) to a variable configuration including 1 to 2 convolutional layers, use a convolutional kernel of size 2, and do not use a pooling layer to capture the dynamic changes of cells in a short time;

[0083] Step 3.3: Based on the Encoder module with self-attention mechanism, single-head attention is adopted, followed by a single-layer feed-forward neural network (FNN). Among them, the Encoder module uses the single-head attention mechanism to process the input information. The implementation of single-head attention is to obtain query, key, and value matrices by linearly transforming the input. Subsequently, the attention mechanism calculates the similarity between the query and the key, assigns different weights to the inputs at different positions, so as to capture the correlation and dependence between the inputs at different positions. After the attention output, a single-layer feed-forward neural network is used to perform further non-linear transformation on the data to enrich the feature expression;

[0084] Step 3.4: Concatenate the output of the last layer of LSTM (i.e., the hidden vector of LSTM) with the flattened output of CNN (i.e., the hidden vector of CNN) as the input of the Encoder module;

[0085] Step 3.5: The output of the Encoder module passes through a Softmax activation function to generate the final binary classification prediction fusion sequence self-attention network model. The Encoder module of the fusion sequence self-attention network model (FSAN) mainly combines the self-attention mechanism and sequence feature extraction technology.

[0086] The expression of the activation function is as follows:

[0087] ;

[0088] Among them, is the number of categories for classifying the input data, is the value corresponding to the rd category, is the value when taking the value of in the th category.

[0089] As a further solution of this embodiment, when training with the Encoder module of FSAN, the Adam optimizer and cross-entropy loss function are used. The preprocessed dataset is divided into a training set, a validation set, and a test set according to the ratio of 6:2:2. All random seeds are set to 202.

[0090] As a further solution of this embodiment, accuracy, precision, recall, and F1 score are used as evaluation indicators for the FSAN model. Precision reflects the proportion of true positives that actually meet the production requirements among the predicted positives, and is an intuitive measure of the model's performance; recall measures the proportion of positive samples identified by the model among all positive samples, which is directly related to the omission rate in the screening process; their harmonic mean, the F1 score, is used as the main evaluation indicator. The experimental results are processed and analyzed. Using F1 as the main evaluation indicator, after more than 5 days of cultivation in the shake flask stage, the F1 score of FSAN stabilizes above 80%, and the best case reaches 94.3%.

[0091] The experimental results of FSAN are shown in Table 2.

[0092] Table 2: Experimental Results of FSAN

[0093] 。 Example 2:

[0094] The present invention also provides a prediction device for cell screening after transfection. Refer to Figure 3 as shown, the prediction device includes:

[0095] An original dataset import unit for obtaining the original dataset for cell screening;

[0096] An original dataset preprocessing unit for preprocessing the original dataset to obtain feature data;

[0097] A cell screening prediction unit for inputting the preprocessed feature data into a fusion prediction model for prediction, obtaining a prediction result, and exporting a prediction result file.

[0098] The original dataset preprocessing unit includes:

[0099] A basic data file module, including a cell screening database and an import module. The cell screening database obtains and stores basic data files, and the import module imports the basic data files into the data preprocessing module; the basic data files are the original dataset;

[0100] A data preprocessing module for padding zeros and marking incomplete time series, performing label balancing using the SMOTE algorithm, normalizing the data, analyzing the correlation of cell screening parameters, and retaining data with a correlation coefficient less than 0.12 as feature data, and outputting the feature data to the cell screening prediction unit.

[0101] The cell screening prediction unit includes:

[0102] A prediction model generation module for training and testing multiple machine learning fusion prediction models using the feature data output by the original dataset preprocessing unit;

[0103] A prediction model storage module that stores multiple trained machine learning fusion prediction models;

[0104] A prediction module that inputs the feature data obtained by preprocessing the original data set into the prediction model, outputs the prediction result, and exports the prediction result file. Embodiment 3:

[0105] The present invention also provides an electronic device, which includes a processor and a memory. The processor is interconnected with the memory. Among them, the memory is used to store a computer program that supports the electronic device to execute the prediction method for cell screening described above. The computer program includes program instructions, and the processor is configured to call the program instructions to execute the prediction method for cell screening described in any one of the above. Embodiment 4:

[0106] The present invention also provides a storage medium that stores a computer program. The computer program includes program instructions, and when the program instructions are executed by a processor, the processor is caused to execute the prediction method for cell screening described in any one of the above.

[0107] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A prediction method for screening cells after transfection, characterized in that: The following steps are involved: Step 1, obtaining cell screening data as an original data set; the original data set includes multiple data, each of which includes a time series with a length of 10 days and six key characteristic indicators; the six key characteristic indicators are viable cell concentration, total cell concentration, total cell number, average particle size, average roundness and agglomeration rate; Step 2: Preprocess the data in the original data set, and use correlation analysis and significance to test the rationality of the data preprocessing strategy to obtain feature data; the specific process of preprocessing the data in the original data set is as follows: Select any two data from the original data set with label 1, use the median interpolation method to standardize the data from the selected two original data sets with label 1, and generate new sample data x marked as 1′ new ; All data in the original data set are standardized to obtain a standardized original data set; Add ±0.2 random perturbations to the standardized data, and each new sample data x new Generate four new sample data x′ new , get the preprocessed data set; The standardized original data set and the preprocessed data set were subjected to Pearson correlation analysis and t-test to obtain feature data, which included the correlation change r between different features and the p-value of the corresponding feature; Step 3: Construct a fusion sequence self-attention network model, input the obtained feature data into the fusion sequence self-attention network model, and obtain the prediction result.

2. The prediction method for post-transfection cell screening according to claim 1, characterized in that: The process of collecting a single piece of data is as follows: ①. In the deep-well plate stage, data collection was performed only once after four days of culture to obtain one piece of data; ② In the shaking bottle stage, one piece of data is collected on the first day, third day, fifth day, seventh day, eighth day, ninth day, tenth day, twelfth day and fourteenth day of the shaking bottle stage, and nine pieces of data are obtained; and cells that do not meet the requirements are screened out according to the growth of cells, and the incomplete time series are filled with 0; ③. Combine one piece of data obtained in the deep well plate stage with nine pieces of data obtained in the shake bottle stage to obtain a time series with a length of 10 days.

3. The prediction method for post-transfection cell screening according to claim 1, characterized in that: The data in the original data set is stored in a CSV file; The results obtained using Pearson correlation analysis and t-test were stored in CSV files.

4. The prediction method for screening transfected cells according to any one of claims 1 to 3, characterized in that: The process of constructing the fusion sequence self-attention network model is as follows: The LSTM neural network model is configured as a two-layer structure, each layer of which contains a variable configuration of 16 to 64 hidden units, and the dropout function is set to 0.5; The convolutional neural network is set to a variable configuration containing 1 to 2 convolutional layers, using a convolution kernel of size 2 and no pooling layer to capture the dynamic changes of cells in a short period of time; The Encoder module based on the self-attention mechanism uses single-head attention, followed by a single-layer feedforward neural network. The Encoder module uses the single-head attention mechanism to process the input information. The implementation method of single-head attention is to obtain the query, key, and value matrix through linear transformation of the input. The attention mechanism assigns different weights to inputs at different positions by calculating the similarity between the query and the key, thereby capturing the correlation and dependency between inputs at different positions. After the attention output, the data is further nonlinearly transformed through a single-layer feedforward neural network to enrich the feature expression. Concatenate the last layer output of LSTM with the flattened output of CNN as the input of Encoder; The output of the encoder passes through a Softmax activation function to produce the final binary classification prediction fusion sequence self-attention network model.

5. The prediction method for post-transfection cell screening according to claim 4, characterized in that: The specific process of inputting the obtained feature data into the fusion sequence self-attention network model for prediction is as follows: The preprocessed data set is divided into training set, validation set and test set in a ratio of 6:2:2, and all random seeds are set to 202; The Encoder module of the fused sequence self-attention network model is used to train the training set, and the Adam optimizer and cross entropy loss function are used when training the training set. The accuracy, precision, recall and F1 score are used as evaluation indicators of the fused sequence self-attention network model to obtain the prediction results.

6. A post-transfection cell screening prediction device for implementing the post-transfection cell screening prediction method according to any one of claims 1 to 5, characterized in that: The prediction device comprises: The original data set is imported into the unit to obtain the original data set of cell screening; The original data set preprocessing unit preprocesses the original data set to obtain feature data; The cell screening prediction unit inputs the preprocessed feature data into the fusion prediction model for prediction, obtains the prediction results, and exports the prediction result file.

7. An electronic device, characterized in that: including interconnected processors and memories; The memory is used to store a computer program that supports the electronic device to execute the prediction method for cell screening, and the computer program includes program instructions; The processor is configured to call the program instructions to execute the cell screening prediction method as described in any one of claims 1-5.

8. A storage medium, characterized in that: The storage medium stores a computer program, which includes program instructions. When the program instructions are executed by a processor, the processor executes the prediction method for cell screening according to any one of claims 1 to 5.

Images