An intelligent method for preparing a mammalian gene edited cell

Abstract

This application discloses a smart method for preparing gene-edited mammalian cells. The method includes: acquiring the target gene sequence information of the mammal to be observed and processing it according to preset screening rules to obtain the target sgRNA sequence; mixing an RNP complex suspension prepared based on the target sgRNA sequence with a suspension of embryonic stem cells to be edited, and performing electroporation transfection using multiple consecutive micropulses to obtain a transfected cell suspension; acquiring raw images of the transfected cell suspension and preprocessing them to obtain single-cell images, and using a trained lightweight convolutional neural network model to obtain the predicted gene-editing efficiency and prediction confidence after 48 hours; if the transfected cell suspension meets the quality control conditions, expanding and culturing the transfected cell suspension to obtain the prepared gene-edited cells. This method can shorten the cell culture cycle, reduce the waste of cell materials, improve the stability of transfection efficiency, and achieve precise preparation of gene-edited cells.

Description

Technical Field

[0001] This application relates to the fields of biotechnology, genetic engineering and artificial intelligence, specifically to a smart method for preparing mammalian gene-edited cells. Background Technology

[0002] Mammalian embryonic stem cells (mESCs) possess pluripotency and multipotent differentiation potential, making them a core biological carrier for gene therapy, protein drug production, disease-resistant breeding, and improved productive traits. They have irreplaceable application value in the fields of biomedicine, agriculture, and animal husbandry. CRISPR-Cas9-mediated gene editing technology provides a highly efficient tool for the targeted modification of mammalian embryonic stem cells. However, these cells are extremely sensitive to transfection parameters, and the maintenance of pluripotency requires stringent conditions. In traditional preparation processes, the cell membrane integrity and suspension impedance of different batches of mammalian embryonic stem cells can vary by 30-50 Ω. The statically preset electroporation parameters used in traditional preparation processes are difficult to adapt to individual differences, hindering the precise preparation of gene-edited cells. Furthermore, this method, which relies on artificial pre-experimental optimization of transfection parameters, cannot respond to fluctuations in cell state, resulting in unstable transfection efficiency. In addition, this method involves a long cell culture cycle and high cost, and the 48-72 hour post-transfection validation period leads to a double waste of cell materials and time. Summary of the Invention

[0003] In view of this, this application provides a method for the intelligent preparation of gene-edited mammalian cells, the method comprising: The target gene sequence information of the mammal under observation is obtained, and the target gene sequence information is processed according to the preset screening rules to obtain the target sgRNA sequence. The off-target risk probability score corresponding to the target sgRNA sequence is less than or equal to the off-target risk probability threshold, and the predicted value of the cleavage efficiency corresponding to the target sgRNA sequence is greater than or equal to the cleavage efficiency threshold. The RNP complex suspension prepared based on the target sgRNA sequence is mixed with the embryonic stem cell suspension to be edited, and electroporation transfection is performed using multiple consecutive micropulses to obtain the transfected cell suspension. Among them, within 10ms after the end of each micropulse application, a real-time state space is constructed based on the real-time state data of the mixed cell suspension. Inter-state vectors are used to input the real-time state vectors into a trained TD3 model to obtain the adjustment amount of the electroporation parameters for the next micropulse. The electroporation parameters of subsequent micropulses are then adjusted according to the adjustment amount. This process continues until all micropulses are applied, resulting in a transfected cell suspension. The original images of the transfected cell suspension are acquired and preprocessed to obtain single-cell images. These single-cell images are then input into a trained lightweight convolutional neural network model to obtain the predicted gene editing efficiency and corresponding prediction confidence after 48 hours. If the transfected cell suspension meets the quality control conditions based on the predicted gene editing efficiency and corresponding prediction confidence, the transfected cell suspension is amplified and cultured to obtain the prepared gene-edited cells.

[0004] Optionally, the real-time state data includes the real-time impedance value, impedance change rate, real-time viability, and the number of currently unapplied micropulses of the mixed cell suspension; a real-time state space vector is constructed based on the real-time cell state data, including: real-time acquisition of the real-time impedance value of the mixed cell suspension, calculation of the impedance change rate based on the real-time impedance value, and simultaneous acquisition of the propidium iodide fluorescence intensity of the mixed cell suspension, calculation of the real-time viability based on the propidium iodide fluorescence intensity; and construction of the real-time state space vector based on the real-time impedance value, impedance change rate, real-time viability, and the number of currently unapplied micropulses.

[0005] Optionally, the electroporation parameters include the voltage magnitude of the micropulse, and the method further includes: during the electroporation transfection of the mixed cell suspension using multiple consecutive micropulses, if the impedance change rate of the mixed cell suspension is detected to be greater than the impedance change rate threshold, the voltage of the current micropulse is reduced by a preset value, and electroporation transfection is continued based on the voltage-adjusted micropulse.

[0006] Optionally, the training process of the TD3 model includes: acquiring historical data from multiple batches of electroporation experiments; the historical data includes the current state data before each micropulse application, the applied action data, the reward value obtained after applying the action data, and the next state data after applying the action data during multiple electroporation experiments; the current state data includes the reference real-time impedance value, the reference impedance change rate, the reference real-time liveness, and the number of unapplied reference micropulses; the action data includes the reference adjustment amount of the electroporation parameters; constructing multiple quadruplet training data sets based on the historical data, each consisting of the current state data, the applied action data, the reward value obtained after application, and the next state data after applying the action data; inputting the next state data from the quadruplet training data into the Actor network to obtain candidate action data; and combining the candidate action data with the quadruplet training data. The next state data from the training data is concatenated and input into the dual Critic network to obtain two first evaluation values. The target evaluation value is determined based on the smaller of the two first evaluation values ​​and the reward value in the quadruplet training data. The current state data and the applied action data from the quadruplet training data are concatenated and input into the dual Critic network to obtain two second evaluation values. The network parameters of the corresponding Critic network are updated using gradient descent based on the mean square error between the two second evaluation values ​​and the target evaluation value. At each preset update interval, the network parameters of the Actor network are updated using gradient descent based on the quadruplet training data and the current Critic network's evaluation value of the current Actor network's output action, until the convergence condition is met, resulting in the trained TD3 model.

[0007] Optionally, the electroporation parameters include the voltage magnitude of the micropulse, and the reward value is calculated based on a balanced reward function. Represented as: ; in, This represents the predicted transfection efficiency. Indicates cell death rate, The voltage value is adjusted accordingly; the transfection efficiency is predicted by the reference impedance change rate and a pre-established regression model relating the impedance change rate to the transfection efficiency; and the cell death rate is calculated based on the reference real-time viability.

[0008] Optionally, the lightweight convolutional neural network model is built based on the MobileNetV3 model, including the feature extraction layer of the MobileNetV3 model, and the original classification head of the output layer in the MobileNetV3 model is replaced with a regression prediction head; single-cell images are input into the trained lightweight convolutional neural network model to obtain the predicted gene editing efficiency and corresponding prediction confidence after 48 hours, including: using the feature extraction layer to extract features from the single-cell image to obtain 128-dimensional features; and using the regression prediction head to predict the feature vector composed of the 128-dimensional features to obtain the predicted gene editing efficiency and corresponding prediction confidence after 48 hours.

[0009] Optionally, the lightweight convolutional neural network model is trained using a composite loss function. Represented as: ; in, The loss in gene editing efficiency prediction is represented by the mean squared error between the predicted gene editing efficiency and the actual gene editing efficiency. The confidence calibration loss is used to calibrate the prediction accuracy of confidence level versus gene editing efficiency. This is the balance coefficient.

[0010] Optionally, the method further includes: after amplifying and culturing the transfected cell suspension, obtaining the actual gene editing efficiency; if the absolute value of the difference between the actual gene editing efficiency and the predicted gene editing efficiency is greater than the gene editing efficiency threshold, fine-tuning the parameters of the lightweight convolutional neural network model online based on the gene editing efficiency threshold to obtain the optimized lightweight convolutional neural network model.

[0011] Optionally, meeting the quality control conditions includes a predicted gene editing efficiency greater than or equal to a gene editing efficiency threshold, and a prediction confidence level greater than or equal to a confidence threshold. The method further includes: when the predicted gene editing efficiency is less than the gene editing efficiency threshold, and the prediction confidence level is greater than or equal to the confidence threshold, adjusting the initial electroporation parameters based on a preset adjustment amount to obtain reference electroporation parameters; performing secondary electroporation transfection on the transfected cell suspension according to the reference electroporation parameters and multiple consecutive micropulses to obtain a cell suspension after secondary electroporation transfection; obtaining the first updated predicted gene editing efficiency and the corresponding first updated prediction confidence level after 48 hours based on image data collected from the cell suspension after secondary electroporation transfection and a trained lightweight convolutional neural network model; and, if the cell suspension after secondary electroporation transfection meets the quality control conditions based on the first updated predicted gene editing efficiency and the corresponding first updated prediction confidence level, expanding and culturing the cell suspension after secondary electroporation transfection to obtain prepared gene-edited cells.

[0012] Optionally, the method further includes: re-acquiring images of the transfected cell suspension when the prediction confidence is less than a confidence threshold; obtaining the second updated predicted gene editing efficiency and the corresponding second updated predicted confidence after 48 hours based on the re-acquisitioned images and a trained lightweight convolutional neural network model; and expanding and culturing the transfected cell suspension to obtain prepared gene-edited cells when the second updated predicted gene editing efficiency and the corresponding second updated predicted confidence meet the quality control conditions.

[0013] In this embodiment, multiple consecutive micropulses are used to electroporate and transfect a mixed cell suspension. During transfection, the adjustment amount of the electroporation parameters for the next micropulse is predicted using a TD3 model and real-time state data of the mixed cell suspension. The electroporation parameters of subsequent micropulses are adjusted according to the adjustment amount, so that the electroporation parameters of the micropulses applied to the mixed cell suspension are dynamically matched with the real-time state data of the mixed cell suspension. This improves the stability of transfection efficiency and enables precise preparation of gene-edited cells. After transfection of the mixed cell suspension, a lightweight convolutional neural network model and preprocessed single-cell images of the original images of the transfected cell suspension are used to predict the predicted gene editing efficiency and corresponding prediction confidence level of the transfected cell suspension after 48 hours. When the predicted gene editing efficiency and corresponding prediction confidence level meet the quality control conditions, the transfected cell suspension is expanded and cultured, shortening the cell culture cycle and avoiding the waste of cell material caused by waiting 48-72 hours after transfection for post-verification in the prior art. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 This is a flowchart illustrating a method for intelligent preparation of mammalian gene-edited cells, as provided in an embodiment of this application. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0016] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0017] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of this application pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have a meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0018] This application provides a smart method for preparing mammalian gene-edited cells, such as... Figure 1 The diagram shown is a flowchart illustrating a smart preparation method for mammalian gene-edited cells provided in this application. The method includes: S101. Obtain the target gene sequence information of the mammal to be observed, process the target gene sequence information based on the preset screening rules, and obtain the target sgRNA sequence.

[0019] In some embodiments, the mammal to be observed is a mammal whose genome requires gene editing or modification of its embryonic stem cells. The target gene sequence information is the complete genome sequence of the target gene of the mammal to be observed, which is related to a specific trait. The specific trait may include disease resistance, high yield, and production of medicinal proteins. The target gene sequence information can be obtained from publicly available bioinformatics databases. This application uses the complete genome sequence of the mouse TLR4 gene (related to disease resistance) as an example, and its sequence information can be obtained from the NCBI database, accession number NM_021297.3.

[0020] In some embodiments, the preset screening rules may include processes such as scanning exon regions of target gene sequence information to generate candidate sgRNA sequences, filtering candidate sgRNA sequences, calling a pre-trained random forest model to score the off-target risk probability and predict the cleavage efficiency of the filtered sgRNA sequences, and screening based on the off-target risk probability score and cleavage efficiency.

[0021] In some embodiments, the off-target risk probability score corresponding to the target sgRNA sequence is less than or equal to the off-target risk probability threshold, and the predicted cleavage efficiency value corresponding to the target sgRNA sequence is greater than or equal to the cleavage efficiency threshold. The off-target risk probability threshold can be 0.1, 0.15, etc., and the predicted cleavage efficiency value can be 70%. In the process of processing target gene sequence information based on preset screening rules, the coding region and exon / intron boundaries of the mouse TLR4 gene (2712 bp in length) provided in FASTA format can be annotated first (this ensures that the CRISPOR open-source algorithm only scans the key regions in the gene that actually perform functions (such as exons), avoiding the design of invalid sgRNAs in intron regions), and a PAM sequence (such as "NGG", "NG", where N is any base) can be specified. This must match the Cas9 protein (such as SpCas9, Cas9-NG) selected in subsequent experiments to ensure consistency between theoretical design and actual operation. The CRISPOR open-source algorithm is then called to automatically scan all possible sgRNA sequences in the exon region, generating 150-200 candidate sgRNA sequences, each containing a 20 bp target sequence and a 3 bp target sequence. PAM sequences were used for initial filtering, while sequences containing more than 6 consecutive homologous bases and GC content less than 40% or greater than 60% were initially filtered out (sequences with design flaws were removed, as these sequences would directly affect the binding stability of sgRNA and DNA); 20 characteristics of the screened sgRNA sequences were extracted, including: GC content (calculated using 20...). The off-target risk probability score and cleavage efficiency prediction value (0~100, higher score indicates stronger in vitro cleavage activity) of each selected sgRNA sequence were determined by considering the following 20 features: G / C base ratio in the target bp sequence, base entropy value, number of off-target sites (calculated by comparing the whole genome with BLAST, counting the total number of sites with ≤3 mismatches), cleavage efficiency score, PAM neighbor base preference (-3 / -2 base type), and whether it contains a Poly-T sequence. These 20 features were then input into a trained random forest model to obtain the off-target risk probability score and cleavage efficiency prediction value (0~100, higher score indicates stronger in vitro cleavage activity) for each selected sgRNA sequence. Finally, a dual criterion of an off-target risk probability score less than or equal to 0.1 and a cleavage efficiency prediction value greater than or equal to 70 was applied. If the number of sgRNA sequences meeting the criteria was greater than or equal to 3, the sgRNA sequence with the highest cleavage efficiency prediction value was selected as the target sgRNA sequence; if there were fewer than 3, the off-target risk probability threshold was adjusted to 0.15 for re-selection.

[0022] S102. The RNP complex suspension prepared based on the target sgRNA sequence is mixed with the embryonic stem cell suspension to be edited, and electroporation transfection is performed using multiple consecutive micropulses to obtain the transfected cell suspension.

[0023] Within 10 ms after each micropulse application, a real-time state space vector is constructed based on the real-time state data of the mixed cell suspension. This real-time state vector is then input into a trained Twin Delayed Deep Deterministic Policy Gradient (TD3) model to obtain the adjustment amount for the electroporation parameters of the next micropulse. The electroporation parameters of subsequent micropulses are then adjusted based on this adjustment amount. This process continues until all micropulses have been applied, resulting in a transfected cell suspension. The electroporation parameters include the voltage magnitude and pulse width of the micropulse.

[0024] In some embodiments, an RNP (ribonucleoprotein) complex suspension can be prepared based on the target sgRNA sequence and Cas9 protein, and a mouse embryonic stem cell suspension can be prepared simultaneously. Then, the RNP complex suspension and the embryonic stem cell suspension are mixed in an appropriate ratio to obtain a mixed cell suspension.

[0025] In preparing the RNP complex suspension, calculate the required volume based on the target sgRNA sequence to Cas9 protein molar ratio of 1:2, add incubation buffer to a total volume of 150 μL, and gently pipette 5-6 times to mix, avoiding the formation of air bubbles. Incubate the prepared mixture in a 37°C water bath for 15 minutes, gently inverting the centrifuge tube once every 5 minutes to promote the formation of a stable RNP complex between the target sgRNA sequence and Cas9 protein. After incubation, immediately store on ice.

[0026] In preparing mouse embryonic stem cell suspension, the old culture medium in the culture dish was discarded, and the cells were washed twice with 2 mL of PBS buffer. After discarding the PBS, 1 mL of 0.25% trypsin-EDTA digestion solution was added, and the mixture was incubated at 37°C for 3 minutes. After observing under a microscope that the cells became rounded and the gaps between them increased, 2 mL of DMEM medium containing 10% fetal bovine serum was immediately added to stop the digestion. The bottom of the culture dish was gently agitated with a pipette to completely detach the cells and form a single-cell suspension. 10 μL of the embryonic stem cell suspension was mixed with 10 μL of trypan blue staining solution (0.4%), and after standing for 3 minutes, it was added to the sample cell of a cell counter. The voltage threshold was set to 80V and the particle size threshold to 5-20 μm. The total number of cells and the number of viable cells were recorded, and the pre-transfection viability was calculated. If the pre-transfection viability was <90%, the cell suspension was centrifuged at 1000 r / min for 5 minutes, the supernatant was discarded, and the cells were resuspended in fresh culture medium and retested. If the result was still unsatisfactory, the cell batch was replaced.

[0027] In some embodiments, after obtaining the mixed cell suspension, the mixed cell suspension can be loaded into an electroporation cuvette, and a single pulse can be split into multiple consecutive micropulses to perform electroporation transfection. Real-time state data is collected within 10ms after the end of each micropulse, a real-time state space vector is constructed based on the real-time state data, and a trained TD3 model is used to predict the adjustment amount of the electroporation parameters for the next micropulse based on the real-time state space vector. The electroporation parameters of subsequent micropulses are then adjusted according to the adjustment amount. The number of micropulses can be 3, 5, 6, etc. For example, if the number of micropulses is 3, the first real-time state data of the mixed cell suspension can be collected within 10ms after the first micropulse is applied. The first real-time state space vector constructed from the first real-time state data is input into the trained TD3 model to obtain the first adjustment amount of the electroporation parameters for the second micropulse. The electroporation parameters of the second micropulse are adjusted based on the first adjustment amount. Then, the second micropulse with the adjusted electroporation parameters is applied. The second real-time state data of the mixed cell suspension is collected within 10ms after the second micropulse is applied. The second real-time state space vector constructed from the second real-time state data is input into the trained TD3 model to obtain the second adjustment amount of the electroporation parameters for the third micropulse. The electroporation parameters of the third micropulse are adjusted based on the second adjustment amount. The transfected cell suspension can be obtained after the third micropulse is applied.

[0028] In some embodiments, during electroporation transfection of the mixed cell suspension using multiple consecutive micropulses, if the impedance change rate of the mixed cell suspension is detected to be greater than an impedance change rate threshold, the voltage of the current micropulse is reduced by a preset value, and electroporation transfection continues based on the voltage-adjusted micropulse. The impedance change rate threshold can be 10. / ms, 15 / ms, the preset value can be any value in the range of [5V, 10V]. When the impedance change rate is greater than the impedance change rate threshold, it indicates that the current risk of cell membrane damage in the mixed cell suspension has increased sharply. At this time, the point puncture parameter of the currently applied micropulse can be lowered, for example, the voltage of the micropulse can be reduced by 5V, 8V or 10V, so as to continue to perform electroporation transfection based on the adjusted micropulse and reduce the probability of cell damage.

[0029] In some embodiments, real-time state data includes the real-time impedance value, impedance change rate, real-time viability, and number of currently unapplied micropulses of the mixed cell suspension. During the process of constructing a real-time state space vector based on the real-time cell state data, the real-time impedance value of the mixed cell suspension can be acquired in real time, the impedance change rate can be calculated based on the real-time impedance value, and the propidium iodide fluorescence intensity of the mixed cell suspension can be acquired simultaneously, and the real-time viability can be calculated based on the propidium iodide fluorescence intensity. The real-time state space vector is constructed based on the real-time impedance value, impedance change rate, real-time viability, and number of currently unapplied micropulses.

[0030] In some embodiments, a platinum electrode sensor can be used to acquire the real-time impedance value of the mixed cell suspension at a frequency of 1 kHz within 10 ms after each micropulse, and the impedance change rate can be calculated based on the real-time impedance value. Simultaneously, fluorescence detection is initiated to acquire the fluorescence intensity of propidium iodide (PI), and the real-time viability is calculated based on the PI fluorescence intensity.

[0031] In some embodiments, the real-time impedance value, impedance change rate, real-time liveness, and number of currently unapplied micropulses can be combined to construct a time-space vector as input to a trained TD3 model to predict the adjustment amount of the electroporation parameters for the next micropulse.

[0032] S103. Acquire the original image of the transfected cell suspension and preprocess it to obtain single-cell images. Input the single-cell images into a trained lightweight convolutional neural network model to obtain the predicted gene editing efficiency and corresponding prediction confidence after 48 hours.

[0033] In some embodiments, the original image of the transfected cell suspension may include a bright-field image, a DAPI channel image, and a GFP channel image acquired from the transfected cell suspension. Preprocessing the original image to obtain a single-cell image includes identifying the cell nucleus boundary in the DAPI channel image, and based on the identified cell nucleus boundary information (such as the center coordinates and size of the cell nucleus), cropping the region where the cell nucleus is located in the bright-field image, the DAPI channel image, and the GFP channel image, and denoising the cropped cell nucleus image to obtain multiple 64×64 pixel single-cell images.

[0034] For example, the transfected cell suspension can be transferred to a confocal culture dish, 2 mL of complete culture medium preheated to 37°C can be added, and the dish can be incubated at 37°C with 5% CO2 for 4 hours. Then, the culture dish is placed on the stage of a confocal microscope, and three fields of view (avoiding edge areas) are selected in different regions of the culture dish to simultaneously acquire bright-field images, DAPI channel images, and GFP channel images. The cell nucleus boundary is identified by the DAPI channel image, and the region where the cell nucleus is located in the bright-field image, DAPI channel image, and GFP channel image is cropped based on the identified cell nucleus boundary information. 20-30 cell nucleus images are cropped from each type of image to obtain about 60-90 cell nucleus images. Finally, the background noise of each cell nucleus image is removed by a Gaussian filtering algorithm to obtain a single-cell image.

[0035] In some embodiments, a lightweight convolutional neural network model can extract multidimensional features associated with gene editing efficiency from single-cell images, such as nucleocytoplasmic ratio, GFP fluorescence distribution entropy, cell roundness index, mean GFP fluorescence intensity, and cell nucleus edge smoothness. Then, based on the multidimensional features, the gene editing efficiency and confidence level of the transfected cell suspension after 48 hours can be predicted.

[0036] S104. If the transfected cell suspension meets the quality control conditions based on the predicted gene editing efficiency and the corresponding prediction confidence level, the transfected cell suspension is expanded and cultured to obtain the prepared gene-edited cells.

[0037] In some embodiments, the quality control conditions for the transfected cell suspension include a predicted gene editing efficiency greater than or equal to a gene editing efficiency threshold and a predicted confidence level greater than or equal to a confidence level threshold. The gene editing efficiency threshold can be any value within the range of [0.1, 0.5], preferably 0.3, and the confidence level threshold can be any value within the range of [0.8, 1). If the predicted gene editing efficiency and corresponding predicted confidence level of the transfected cell suspension after 48 hours meet the quality control conditions, it indicates that the quality of the transfected cell suspension is qualified (meets the quality control conditions), and extended culture can continue: the qualified transfected cell suspension is seeded into a multi-well culture plate, an appropriate amount of culture medium is added to each well, and the plate is continuously cultured under culture conditions suitable for cell growth; fresh culture medium is replaced regularly to maintain cell viability and avoid nutrient depletion or accumulation of metabolic waste; cell morphology and confluence changes are observed daily, and if signs of cell differentiation appear, the concentration of the relevant reagents for maintaining pluripotency is adjusted in a timely manner to ensure the stability of the undifferentiated cell state. After the cells are cultured to a suitable level of confluence, they are detached from the cell wall using conventional cell digestion methods. Serum-containing culture medium is added to stop the digestion, and the cells are gently pipetted to form a single-cell suspension, which is then collected into centrifuge tubes to obtain the prepared gene-edited cells.

[0038] In some embodiments, if the predicted gene editing efficiency is less than the gene editing efficiency threshold and the prediction confidence is greater than or equal to the confidence threshold (the transfected cell suspension does not meet the quality control conditions), it indicates that the lightweight convolutional neural network model determines the gene editing efficiency to be low. In this case, the initial electroporation parameters can be adjusted based on a preset adjustment amount to obtain reference electroporation parameters. A second electroporation transfection is performed on the transfected cell suspension according to the reference electroporation parameters and multiple consecutive micropulses to obtain a cell suspension after second electroporation transfection. Based on the image data collected from the cell suspension after second electroporation transfection and the trained lightweight convolutional neural network model, the first updated predicted gene editing efficiency and the corresponding first updated prediction confidence are obtained after 48 hours. If the cell suspension after second electroporation transfection meets the quality control conditions based on the first updated predicted gene editing efficiency and the corresponding first updated prediction confidence, the cell suspension after second electroporation transfection is amplified and cultured to obtain prepared gene-edited cells. The initial electroporation parameter refers to the voltage magnitude of the first micropulse applied during electroporation transfection, and the preset adjustment amount can be any value within [-10V, +10V]. It should be noted that during this process, if the predicted gene editing efficiency of the first update is determined to be less than the gene editing efficiency threshold, and the confidence level of the first update prediction is greater than or equal to the confidence threshold (meaning the cell suspension after secondary electroporation transfection still does not meet the quality control conditions), the initial electroporation parameter can be adjusted again based on the preset adjustment amount, and a second electroporation transfection can be performed until it is determined that the cell suspension after multiple transfections meets the quality control conditions. Then, amplification culture is performed to complete the preparation of gene-edited cells. This process can improve gene editing efficiency.

[0039] In some embodiments, if the prediction confidence is less than the confidence threshold, it may indicate poor image quality or abnormal cell state. In this case, images of the transfected cell suspension can be re-acquired. Based on the re-acquired images and the trained lightweight convolutional neural network model, a second updated predicted gene editing efficiency and corresponding second updated prediction confidence are obtained after 48 hours. If the second updated predicted gene editing efficiency and corresponding second updated prediction confidence meet the quality control conditions, the transfected cell suspension is amplified and cultured to obtain the prepared gene-edited cells. It should be noted that during this process, if the second updated prediction confidence is determined to be less than the confidence threshold, images of the transfected cell suspension can be re-acquired. Based on the re-acquired images and the trained lightweight convolutional neural network model, a new predicted gene editing efficiency and prediction confidence are obtained. This process is repeated until the prediction confidence output by the model is greater than or equal to the confidence threshold, and the gene editing efficiency output by the model is greater than or equal to the gene editing efficiency threshold. Then, amplification and culture are performed to complete the preparation of gene-edited cells. This process can avoid the risk of misjudgment caused by imaging anomalies, ensure the quality of single-cell images input to the lightweight convolutional neural network model, thereby improving the robustness and reliability of the model's prediction results and avoiding the situation where the preparation process is erroneously triggered by a single low-quality sampling.

[0040] It should be noted that the method of performing electroporation transfection again when the predicted gene editing efficiency is less than the gene editing efficiency threshold and the prediction confidence is greater than or equal to the confidence threshold, and the method of re-acquiring images of the transfected cell suspension for prediction when the prediction confidence is less than the confidence threshold, are not independent. The two methods can be performed in combination.

[0041] In this embodiment, the target gene sequence information of the mammal to be observed is obtained, and the target gene sequence information is processed based on a preset screening rule to obtain the target sgRNA sequence. An RNP complex suspension prepared based on the target sgRNA sequence is mixed with an embryonic stem cell suspension to be edited, and electroporation transfection is performed using multiple consecutive micropulses. During transfection, the adjustment amount of the electroporation parameters for the next micropulse is predicted using the TD3 model and real-time state data of the mixed cell suspension. The electroporation parameters of subsequent micropulses are adjusted according to the adjustment amount, so that the electroporation parameters of the micropulses applied to the mixed cell suspension are consistent with the real-time state of the mixed cell suspension. Dynamic data matching improves the stability of transfection efficiency and enables precise preparation of gene-edited cells. After transfection with the mixed cell suspension, the original image of the transfected cell suspension is acquired and preprocessed to obtain single-cell images. These single-cell images are then input into a trained lightweight convolutional neural network model to obtain the predicted gene editing efficiency and corresponding prediction confidence after 48 hours. If the predicted gene editing efficiency and corresponding prediction confidence meet the quality control conditions, the transfected cell suspension is expanded and cultured, shortening the cell culture cycle and avoiding the waste of cell material caused by waiting 48-72 hours after transfection for post-verification in existing technologies.

[0042] In some embodiments of this application, after amplifying and culturing the transfected cell suspension, the actual gene editing efficiency can be obtained; if the absolute value of the difference between the actual gene editing efficiency and the predicted gene editing efficiency is greater than the gene editing efficiency threshold, the parameters of the lightweight convolutional neural network model are fine-tuned online based on the gene editing efficiency threshold to obtain the optimized lightweight convolutional neural network model.

[0043] In some embodiments, methods such as decomposition algorithms based on Sanger sequencing (e.g., TIDE / ICE), targeted amplicon deep sequencing based on high-throughput sequencing (e.g., CRISPResso2 analysis), and fluorescence reporter system detection based on flow cytometry (e.g., GFP positivity rate) can be used to quantitatively analyze changes in the DNA sequence or functional expression products of the edited cell population in the transfected cell suspension to determine the true gene editing efficiency.

[0044] In some embodiments, the gene editing efficiency threshold can be set to 5%. When the absolute value of the difference between the actual gene editing efficiency and the predicted gene editing efficiency is greater than the gene editing efficiency threshold, the parameters of the lightweight convolutional neural network model are fine-tuned based on this threshold. This allows for focused optimization of the network parameters in the regression prediction head (output layer) of the lightweight convolutional neural network model, avoiding over-adjustment that could lead to a decrease in the model's generalization ability. A dynamic learning rate (initial learning rate 0.001, gradually decreasing to 0.0001 as the error decreases) is used to balance adjustment speed and stability. Through continuous data feedback and network parameter fine-tuning, the prediction error of the subsequent lightweight convolutional neural network model is reduced to within 5%. This process improves the online prediction accuracy of the lightweight convolutional neural network model through a closed-loop feedback mechanism.

[0045] In some embodiments of this application, the TD3 model includes an Actor network and a dual Critic network. During the training of the TD3 model, the Actor network outputs predicted action data (adjustments to electroporation parameters), and the dual Critic networks output evaluation values ​​Q1 and Q2 for the action data output by the Actor network, respectively. Q1 is the primary evaluation network, which learns the correlation between impedance changes, voltage adjustments, and reward feedback based on historical data. Q2 is the auxiliary evaluation correction network, which independently learns the same task and reduces the evaluation bias of a single network through cross-validation. For example, when the rate of impedance change changes abruptly during electroporation (increased risk of cell membrane damage), Q1 may overestimate the value of "reducing the voltage by 5V" due to local sample noise, while Q2, with independent weights, can provide a more conservative evaluation. The minimum of the two values ​​more accurately reflects the long-term benefits of the action data, suppresses overestimation, and ensures the rationality of the decision.

[0046] In some embodiments, the training process of the TD3 model includes: acquiring historical data from multiple batches of electroporation experiments; constructing multiple quadruplet training data sets based on the historical data, each consisting of current state data, applied action data, reward value obtained after application, and next state data after applying the action data; inputting the next state data from the quadruplet training data into the Actor network to obtain candidate action data; concatenating the candidate action data and the next state data from the quadruplet training data, and inputting them into a dual Critic network to obtain two first evaluation values, and determining the target evaluation value based on the smaller of the two first evaluation values ​​and the reward value from the quadruplet training data. The value is determined by concatenating the current state data and applied action data from the quadruplet training data and inputting them into a dual Critic network to obtain two second evaluation values. Based on the mean square error between the two second evaluation values ​​and the target evaluation value, the network parameters of the corresponding Critic network are updated using gradient descent. At each preset update interval, based on the quadruplet training data and the current Critic network's evaluation value of the current Actor network's output action, the network parameters of the Actor network are updated using gradient descent until the convergence condition is met (such as the loss function value reaching a preset threshold or reaching a preset number of training rounds), resulting in a trained TD3 model.

[0047] In some embodiments, historical data includes the current state data before each micropulse is applied during multiple electroporation experiments, the applied action data, the reward value obtained after applying the action data, and the next state data after applying the action data; the current state data includes the reference real-time impedance value, the reference impedance change rate, the reference real-time liveness, and the number of reference micropulses not applied; the action data includes the reference adjustment amounts of the electroporation parameters (voltage value, pulse width). Multiple quadruplet training data sets constructed based on historical data can be stored in an experience replay buffer, from which small batches of data can be randomly sampled during training to break data correlation.

[0048] In some embodiments, the network parameters of the Actor network and Critic network can be initialized using a Xavier normal distribution, with the bias term initialized to 0 to ensure stable output in the early stages of training. The learning rate is set to 0.001 to avoid training oscillations; the discount factor is 0.99 to emphasize long-term rewards and ensure overall optimization of multiple micro-pulse sequences, rather than local optima for a single pulse. The preset update epochs can be 10 or 12, with the Critic network parameters updated every 10 or 12 epochs to slow down evaluation fluctuations and improve training stability. To adapt to the voltage adjustment precision of the electroporation device and avoid excessive adjustment that could cause cell membrane rupture, the adjustment amount of the micro-pulse voltage value is [specified]. To balance pulse energy and cell repair time, and to avoid excessive pulse width fluctuations affecting transfection consistency, the pulse width adjustment amount is: .

[0049] In some embodiments, the target evaluation value can be calculated using the following formula (1). : (1); in, This represents the reward value calculated using the balanced reward function. This represents the discount factor, set to 0.99. This indicates the first evaluation value. This indicates the second evaluation value.

[0050] In some embodiments, a balanced reward function It can be expressed by formula (2): (2); in, This represents the predicted transfection efficiency. Indicates cell death rate, The voltage adjustment is indicated by the voltage value. The predicted transfection efficiency is obtained through the reference impedance change rate and a pre-established regression model relating the impedance change rate to transfection efficiency. The cell death rate is calculated based on the reference real-time viability. The reference impedance change rate and cell death rate are data obtained from historical experiments. The regression model relating the impedance change rate to transfection efficiency can be a pre-trained model based on historical data (impedance change rate and transfection efficiency measured in historical experiments).

[0051] Understandably, in a balanced reward function, This can ensure the effectiveness of gene editing; It can constrain parameter adjustments to prevent cell damage and avoid sacrificing survival in pursuit of efficiency; It can control the adjustment range of the voltage magnitude of the micropulse, ensuring the stability of electroporation parameters and avoiding fluctuations in transfection effect caused by drastic fluctuations.

[0052] In some embodiments of this application, the lightweight convolutional neural network model is constructed based on the MobileNetV3 model, including the feature extraction layer of the MobileNetV3 model, and the original classification head of the output layer in the MobileNetV3 model is replaced with a regression prediction head; the step S103 of "inputting the single-cell image into the trained lightweight convolutional neural network model to obtain the predicted gene editing efficiency and the corresponding prediction confidence after 48 hours" includes: using the feature extraction layer to extract features from the single-cell image to obtain 128-dimensional features; using the regression prediction head to predict the vector composed of the 128-dimensional features to obtain the predicted gene editing efficiency and the corresponding prediction confidence after 48 hours.

[0053] In some embodiments, the feature extraction layer can extract 128-dimensional features related to gene editing efficiency from single-cell images, such as nucleocytoplasmic ratio, GFP fluorescence distribution entropy, cell roundness index, average GFP fluorescence intensity, and cell nucleus edge smoothness, to achieve accurate capture of features related to editing effect, such as RNP nuclear entry efficiency and cell morphological integrity.

[0054] In some embodiments, during the training of the lightweight convolutional neural network model, preprocessed single-cell images from multiple batches of gene editing experiments and paired data of actual editing efficiency over 48 hours are collected. Key feature labels such as nucleocytoplasmic ratio and fluorescence distribution entropy are labeled to construct a labeled training set. The lightweight convolutional neural network model is first pre-trained using a general cell image dataset, and then the network parameters are fine-tuned based on labeled data from gene editing scenarios, focusing on optimizing the sensitivity of the feature extraction layer to features related to editing efficiency. A composite loss function is used, with the difference between the predicted editing efficiency and the actual editing efficiency as the optimization objective. A confidence loss term is also introduced (penalizing low-confidence prediction results) to improve prediction accuracy and reliability. The network parameters are iteratively updated using stochastic gradient descent until the error between the model's predicted value and the actual value converges to a preset threshold, and the trained model weight file is output. It should be noted that during the update of network parameters, the original convolutional layer parameters of the MobileNetV3 model (convolutional kernel size, stride, etc.) are retained, and only the network parameters of the regression prediction head are adjusted to adapt to the regression analysis of 128-dimensional feature vectors.

[0055] In some embodiments, the composite loss function used during the training of the lightweight convolutional neural network model This can be expressed by formula (3): (3); in, The loss in gene editing efficiency prediction is represented by the mean squared error between the predicted gene editing efficiency and the actual gene editing efficiency. The confidence calibration loss is used to calibrate the prediction accuracy of confidence level versus gene editing efficiency. This is a balancing coefficient. Calibrating the confidence level and the prediction accuracy of gene editing efficiency allows the prediction accuracy of confidence level and gene editing efficiency to be matched, penalizing high-confidence false predictions or low-confidence correct predictions.

[0056] In some embodiments, if the mean square error between the predicted gene editing efficiency and the actual gene editing efficiency is greater than a preset threshold (indicating a prediction error). If the mean square error between the predicted gene editing efficiency and the actual gene editing efficiency is less than or equal to the preset threshold (indicating that the prediction is correct). , This is the confidence score output by the lightweight convolutional neural network model. Thus, when the lightweight convolutional neural network model makes a wrong prediction, it is penalized by lowering its confidence score; when it makes a correct prediction, it is penalized by higher confidence scores. Ultimately, the lightweight convolutional neural network model learns to only give high confidence scores when its predictions are accurate.

[0057] It is understandable that training a lightweight convolutional neural network model with a composite loss function that includes gene editing efficiency prediction loss and confidence calibration loss can improve the prediction accuracy of the lightweight convolutional neural network model for gene editing efficiency and confidence, and ensure the fit between gene editing efficiency and confidence.

[0058] To verify the irreplaceable role of the TD3 algorithm used in the intelligent preparation method of mammalian gene-edited cells provided in this application in terms of multi-objective balance, real-time responsiveness, batch adaptability and stability, experiments were conducted to compare the regulatory effects of the TD3 algorithm used in this application with existing PID control, random search, Bayesian optimization and traditional static parameter methods in the electroporation transfection of mammalian embryonic stem cells (mESC).

[0059] The experimental conditions were standardized as follows: mouse TLR4 gene (disease resistance related) modified mESCs were selected, and 10 different batches with initial impedance differences of 30-50Ω were chosen; a Venid electroporator of the same model was used, with 5 micropulses spaced 500ms apart, voltage adjustable from 100-140V and pulse width adjustable from 10-20ms; a balanced reward function was used as the evaluation system to balance transfection efficiency, cell viability and parameter stability; cell viability was detected by trypan blue counting 30 minutes after transfection, and transfection efficiency was detected by T7E1 enzyme digestion. The response time of parameter adjustment and batch standard deviation were recorded simultaneously to ensure the comparability and reliability of the experimental results. The experimental results are shown in Table 1.

[0060] Table 1. Comparison of evaluation indicators for experimental results from different methods As can be seen from Table 1, the intelligent preparation method for mammalian gene-edited cells based on the TD3 algorithm outperforms other methods in terms of cell viability, standard deviation of transfection efficiency, parameter adjustment response time, and reward value. This indicates that the intelligent preparation method for mammalian gene-edited cells based on the TD3 algorithm provided in this application can simultaneously meet the requirements of real-time response, multi-objective balance, and resistance to batch-to-batch differences in mESC electroporation transfection.

[0061] It should be noted that, depending on the implementation needs, the steps described in the embodiments of this application can be broken down into more steps, or two or more steps or parts of the steps can be combined into new steps to achieve the purpose of the embodiments of this application.

[0062] The methods described in the embodiments of this application can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code downloaded over a network that is originally stored in a remote recording medium or a non-transitory machine-readable medium and will be stored in a local recording medium. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for executing the methods shown herein.

[0063] Those skilled in the art will recognize that the units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this application.

[0064] The above embodiments are only used to illustrate the embodiments of this application, and are not intended to limit the embodiments of this application. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the embodiments of this application. Therefore, all equivalent technical solutions also fall within the scope of the embodiments of this application, and the patent protection scope of the embodiments of this application should be defined by the claims.

Claims

1. A smart method for preparing mammalian gene-edited cells, characterized in that, include: The target gene sequence information of the mammal to be observed is obtained, and the target gene sequence information is processed based on the preset screening rules to obtain the target sgRNA sequence; The off-target risk probability score corresponding to the target sgRNA sequence is less than or equal to the off-target risk probability threshold, and the predicted cleavage efficiency value corresponding to the target sgRNA sequence is greater than or equal to the cleavage efficiency threshold. A suspension of the RNP complex based on the target sgRNA sequence was mixed with a suspension of embryonic stem cells to be edited, and electroporation transfection was performed using multiple consecutive micropulses to obtain a transfected cell suspension. Within 10 ms after the end of each micropulse application, a real-time state space vector was constructed based on the real-time state data of the mixed cell suspension. The real-time state vector was input into a trained TD3 model to obtain the adjustment amount of the electroporation parameters for the next micropulse, and the electroporation parameters of subsequent micropulses were adjusted according to the adjustment amount. This process was repeated until all micropulses were applied to obtain the transfected cell suspension. Raw images of transfected cell suspensions were acquired and preprocessed to obtain single-cell images. The single-cell images were then input into a trained lightweight convolutional neural network model to obtain the predicted gene editing efficiency and corresponding prediction confidence after 48 hours. After determining that the transfected cell suspension meets the quality control conditions based on the predicted gene editing efficiency and the corresponding prediction confidence level, the transfected cell suspension is expanded and cultured to obtain the prepared gene-edited cells.

2. The method according to claim 1, characterized in that, Real-time status data includes the real-time impedance value of the mixed cell suspension, the rate of change of impedance, the real-time viability, and the number of micropulses not currently applied; A real-time state space vector is constructed based on the real-time state data of the cells, including: Real-time impedance values ​​of the mixed cell suspension are acquired, impedance change rate is calculated based on the real-time impedance values, and propidium iodide fluorescence intensity of the mixed cell suspension is acquired simultaneously, and real-time viability is calculated based on propidium iodide fluorescence intensity. A real-time state space vector is constructed based on the real-time impedance value, impedance change rate, real-time liveness, and the number of currently unapplied micropulses.

3. The method according to claim 2, characterized in that, The electroporation parameters include the voltage magnitude of the micropulse, and the method further includes: During the electroporation transfection of the mixed cell suspension using multiple consecutive micropulses, if the impedance change rate of the mixed cell suspension is detected to be greater than the impedance change rate threshold, the voltage of the current micropulse is reduced by a preset value, and the electroporation transfection continues based on the voltage-adjusted micropulse.

4. The method according to claim 2, characterized in that, The training process of the TD3 model includes: Acquire historical data from multiple batches of electroporation experiments; historical data includes the current state data before each micropulse is applied, the action data after the action data is applied, the reward value obtained after the action data is applied, and the next state data after the action data is applied during multiple electroporation experiments; current state data includes the reference real-time impedance value, the reference impedance change rate, the reference real-time viability, and the number of reference micropulses not applied; action data includes the reference adjustment amount of electroporation parameters; Based on historical data, construct multiple quadruples of training data, consisting of current state data, applied action data, reward value obtained after application, and next state data after applying action data; The next state data from the quadruples training data is input into the Actor network to obtain candidate action data; The candidate action data and the next state data in the quadruple training data are concatenated and input into the dual Critic network to obtain two first evaluation values. The target evaluation value is determined based on the smaller of the two first evaluation values ​​and the reward value in the quadruple training data. The current state data and applied action data in the quadruplet training data are concatenated and input into the dual Critic network to obtain two second evaluation values. Based on the mean square error between the two second evaluation values ​​and the target evaluation value, the network parameters of the corresponding Critic network are updated by gradient descent. At each preset number of update rounds, based on the quadruplet training data and the current Critic network's evaluation value of the current Actor network's output action, the network parameters of the Actor network are updated using gradient descent until the convergence condition is met, resulting in a trained TD3 model.

5. The method according to claim 4, characterized in that, Electroporation parameters include the voltage magnitude of the micropulse, and the reward value is calculated based on a balanced reward function. Represented as: ； in, This represents the predicted transfection efficiency. Indicates cell death rate, The voltage value is adjusted accordingly; the transfection efficiency is predicted by the reference impedance change rate and a pre-established regression model relating the impedance change rate to the transfection efficiency; and the cell death rate is calculated based on the reference real-time viability.

6. The method according to claim 1, characterized in that, The lightweight convolutional neural network model is built on the MobileNetV3 model, including the feature extraction layer of the MobileNetV3 model, and the original classification head of the output layer of the MobileNetV3 model is replaced with a regression prediction head. Single-cell images are input into a trained lightweight convolutional neural network model to obtain the predicted gene editing efficiency and corresponding prediction confidence after 48 hours, including: Feature extraction layers were used to extract features from single-cell images, resulting in 128-dimensional features. The regression prediction head is used to predict the feature vector composed of 128 features, and the predicted gene editing efficiency and corresponding prediction confidence are obtained after 48 hours.

7. The method according to claim 1, characterized in that, The lightweight convolutional neural network model is trained using a composite loss function. Represented as: in, The loss in gene editing efficiency prediction is represented by the mean squared error between the predicted gene editing efficiency and the actual gene editing efficiency. The confidence calibration loss is used to calibrate the prediction accuracy of confidence level versus gene editing efficiency. This is the balance coefficient.

8. The method according to any one of claims 1 to 7, characterized in that, The method further includes: After amplifying and culturing the transfected cell suspension, the actual gene editing efficiency was obtained; If the absolute value of the difference between the actual gene editing efficiency and the predicted gene editing efficiency is greater than the gene editing efficiency threshold, the parameters of the lightweight convolutional neural network model are fine-tuned online based on the gene editing efficiency threshold to obtain the optimized lightweight convolutional neural network model.

9. The method according to claim 1, characterized in that, Meeting the quality control conditions includes predicting gene editing efficiency greater than or equal to a gene editing efficiency threshold, and predicting confidence level greater than or equal to a confidence threshold; the method further includes: If the predicted gene editing efficiency is less than the gene editing efficiency threshold and the prediction confidence is greater than or equal to the confidence threshold, the initial electroporation parameters are adjusted based on the preset adjustment amount to obtain the reference electroporation parameters. A second electroporation transfection was performed on the transfected cell suspension based on reference electroporation parameters and multiple consecutive micropulses to obtain a cell suspension after second electroporation transfection. Based on image data collected from cell suspensions after secondary electroporation transfection and a trained lightweight convolutional neural network model, the first update prediction gene editing efficiency and the corresponding first update prediction confidence were obtained after 48 hours. If the cell suspension after secondary electroporation transfection meets the quality control conditions based on the first update prediction gene editing efficiency and the corresponding first update prediction confidence, the cell suspension after secondary electroporation transfection is expanded and cultured to obtain the prepared gene-edited cells.

10. The method according to claim 9, characterized in that, The method further includes: If the prediction confidence level is less than the confidence threshold, re-acquire images of the transfected cell suspension; Based on the reacquired images and the trained lightweight convolutional neural network model, the second update prediction gene editing efficiency and the corresponding second update prediction confidence were obtained after 48 hours. Under the condition that the second update predicted gene editing efficiency and the corresponding second update predicted confidence level meet the quality control conditions, the transfected cell suspension is expanded and cultured to obtain the prepared gene-edited cells.

Images