Preliminary Study On Image Analysis Techniques During The Growth Of Pichia Pastoris

During the fermentation process of microorganisms,variables such as temperature,p H,dissolved oxygen,substrate and product concentrations,respiratory entropy,and biomass are mainly monitored.According to the changes of the monitored values,the fermentation strategy can be adjusted in time to ensure the output and quality of the product.At present,the monitoring of microbial cells is mainly to characterize the biomass by measuring OD₆₀₀ value,and less attention is paid to the morphology.Changes in microbial cell morphology are associated with changes in cell physiology,often related to productivity,and can be influenced by process conditions.Pichia pastoris is one of the most widely used yeast species in the pharmaceutical and biotechnology industries,grown in culture and induced with different carbon sources.In this study,microscopic images of yeast cells under different magnification conditions obtained during the growth process were used to develop a method for classifying microscopic image data based on deep learning to obtain relevant morphological information.In order to reduce the difficulty of image recognition caused by yeast cell aggregation,the bacterial suspension was dispersed.The main research contents and conclusions are as follows:（1）Different biomass detection methods were measured and compared during the growth of Pichia pastoris.In this study,Pichia pastoris was first cultured with glucose as the carbon source,and the OD₆₀₀ value and dry weight of cells were measured at different times,and the viable cells were counted by the dilution plate method.Comparing the plate cell counting results,it was found that the OD₆₀₀ value did not increase linearly with it.There is no linear correlation between the dilution factor and the OD₆₀₀ value in the logarithmic phase and the stable phase,but a power function relationship.Glucose,glycerol,and methanol were used as carbon sources for growth and culture,and their microscopic images were obtained and observed.（2）To establish a yeast cell dispersion treatment method.The bacterial suspension was washed with a buffer solution with a p H of 6 and an ion concentration of 0.07 mol·L^-1.Tween20,EDTA and ultrasonic treatment were used to reduce cell aggregation.The best dispersion effect was 0.6%Tween 20,8 mmol·L^-1EDTA and 15 s ultrasonic treatment.Orthogonal experiments showed that the optimal conditions were 0.4%Tween 20,8 mmol·L^-1EDTA,and ultrasonic time of 10 s.Under these conditions,the dispersion index of yeast cells was 0.21.（3）YOLOv3 is used to establish a classification model for microscopic images magnified400 times.During the stationary phase of yeast cell growth,the aggregation of cell components was found to be reduced,especially in the stationary phase of growth with methanol as the carbon source.First,the microscopic images of yeast cells grown in methanol to the stationary phase were processed,and the YOLOv3 network was used to detect targets on the images magnified 400 times.A total of 430 labeled images were made as a data set（2212 single cells,3259 budding cells,and 3296 multiple cells）.Perform 200 optimization trainings in the model.The final precision was 85.40%.Then input the yeast cell images at different times into the optimized model,and extract the number and proportion of different categories at different times.（4）Build a BAM-Res Net18 model to classify microscopic images magnified 1000 times.Replace the VGGNet in the original B-CNN model with a set of deep residual network Res Net18 as a feature extractor to improve the extraction of fine-grained features,and add a BAM attention mechanism to each layer to improve classification accuracy.The training set and test set are divided according to the principle of 8:2.Stage one is sorting by single cell,budding cell,and multiple cells.Stage two is to subdivide budding cells and multiple cells on the basis of stage one.The number of training rounds in stage one is 50 times,and the final classification accuracy is 97.83%.The number of training rounds in stage two classification is80 times,and the final classification accuracy is 92.31%.Compared with the original deep learning algorithm,the classification accuracy is improved.