Study On Cell Line Stability Prediction Based On Cell Image

Cell line stability refers to the relative stable state of a cell line expressing a certain target product during passaging and cultivation.Developing stable cell lines is a key technology in the field of biopharmaceuticals,and the stability of cell lines needs to be evaluated before they can be used for large-scale production.Traditional evaluation methods require continuous passaging and cultivation of cell lines for about 90 days to determine their stability.To address the limitations of traditional evaluation methods,this study applies traditional machine learning methods and deep learning techniques to microscopic cell image analysis,and conducts in-depth research on the evaluation and prediction of cell line stability,proposing a method based on cell microscopic image to predict cell line stability.The main research content and contributions of this thesis include the following two aspects:(1)Research on CHO cell segmentation based on channel attention mechanism and dilated convolutionThis paper proposes a semantic segmentation network,RDCA U-Net,based on channel attention mechanism and dilated convolution,to improve cell segmentation accuracy.RDCA U-Net is based on the U-Net network,replacing the convolution module in the original encoder part with a cascade of dilated convolution modules,and introducing residual connections and augmented attention modules,which can enhance the multi-scale feature extraction and fusion ability and have a stronger perception ability for cells of different sizes and shapes.On a CHO cell dataset constructed by the author,RDCA U-Net achieves the highest pixel-level evaluation metrics of PA,JSI,and DSC,with values of 88.98%,82.03%,and 90.13%,respectively.The segmentation performance is better than that of classical semantic segmentation networks and provides more accurate data support for subsequent stability prediction studies.(2)Research on prediction of cell line stability based on cell morphological features and machine learningThis study proposes a method that combines cell geometry features,texture features,and apoptosis parameters to construct a cell stability feature vector,which is used to describe the stability state of individual cells in a cell population and predict the stability trend of the cell line.Classic geometric parameters such as cell perimeter and circularity are selected as geometric feature parameters,while five second-order statistics generated by GLCM,including heterogeneity,angular second moment,contrast,correlation,and inverse difference moment,are used as texture feature parameters.Based on the ResNet34 classification model,a corresponding apoptosis probability triplet is generated for each input cell,which is used to describe the physiological activity state of the cell.The cell stability feature vector constructed by the proposed method is used as input for machine learning algorithms such as SVM,RF,and LR,and the model is tested using unknown data to verify the effectiveness of the proposed method.SVM achieved the best performance with an accuracy of 77.22%,indicating that the method proposed in this study for constructing cell stability feature vectors is effective in describing the stability state of cells.The proposed method in this study can predict the stability of cell lines by evaluating the stability of individual cells,and can detect the instability trend of cell lines 2～3generations earlier.This provides guidance and recommendations for researchers in cell culture,and can assist in improving the stability and quality of cell lines used in biopharmaceutical production.