Design And Realization Of A Method For Classifying Multiplex Digital PCR Droplets Based On Clustering

Multiplex digital PCR,received widespread attention in recent year,is a nucleic molecule quantification technology.It can detect the quantity of multiple target nucleic acids at one time in an absolute quantitative manner.During the process of multiplex digital PCR,dispersing the sample into a great number of droplets.These droplets show different fluorescence values after PCR.Classifying droplets by fluorescence values is an important basis to caculate the quantity of nucleic.At present,softwares of the digital PCR systems applied maturely on the market can only classify the multiplex PCR droplets via threshold setted by professionals,which is time-consuming,laborious and subjective.The research content of this thesis was classifying droplets of multiplex digital PCR.Based on the clustering algorithms in unsupervised learning and related theories of data mining,this thesis conducted actual investigation and built a mathematical model of the digital PCR system;then designed a method for classifying droplets of multiplex digital PCR.In order to achieve a better effect,this thesis proposed a variety of improved algorithms.The main work was as following:(1)The grid regional DBSCAN clustering algorithm was proposed.Firstly,gridmapping was introduced.Data was divided into grids and grids were used as the basic units of clustering,which greatly reduced the time complexity of the algorithm.Secondly,The concept of grid weights and undirected connectivity was proposed.Using the relative of distances and densities between grids would finish clustering in grid space quickly.It had been proved by tests that the improved clustering algorithm had a better performance in some evaluation standards.(2)This thesis designed the smooth window algorithm,a boundary detection and matching algorithm based on Sobel operator.Firstly,the smooth window algorithm was a binary classifier.After reconstructed with window neighborhood information,grids would be divided into normal grids and abnormal grids according to whether there was a sudden change in density.Secondly,the Sobel operator in the image segmentation field was introduced to detect the grids of cluster boundary and these boundary points would be matched to clusters based on the Gaussian Mixture Model.Empirical researches showed that the above algorithms had high accuracies and they could effectively improve the performance of grid regional DBSCAN clustering in the case of complex data distribution.(3)Based on the above theoretical innovations,a method for classifying multiplex digital PCR droplets was designed.Firstly,data was preprocessed to obtain real fluorescence values.Secondly,anomalies and boundary were detected.Thirdly,the grid regional DBSCAN algorithm was used for preliminary classification.Finally,clusters would be combined based on the characteristics of multiplex digital PCR and the results of classifying droplets could be gotten.The experiments built on the digital PCR detection system of Vibrio Pararhaemolyticus had proved that the proposed method could not only accurately classify droplets with a lower time cost,but also played well on a variety of digital PCR applications.(4)Data analysis system for multiplex digital PCR was realized.The system,based on MVC design model,worked as a desktop application(.exe).This system could efficiently preprocess,classify droplets and count results for the samples.Besides,the results would be displayed to users by visual interfaces and the files could be storaged persistently.In summary,it was an automatic,fast and robust method to classify the droplets of multiplex digital PCR,which would play a certain promotion and positive role in the clinical application of multiplex digital PCR technology.