MALDI-TOF Mass Spectra Classification Based On Convolutional Neural Networks And Denoising Autoencoder

Microbial identification is very important for many industries,because microbes are closely related to human life activities.For example,in the medical field,accurate and rapid microbial identification can help disease diagnosis and treatment.Matrix-assisted laser desorption/ionization time-of-flight(MALDI-TOF)mass spectrometry is the most important technological change and breakthrough in the field of microbial identification.The traditional identification and classification of MALDI-TOF mass spectra are often divided into two steps:automatic feature extraction and classification,and the former relies too much on prior knowledge of experts,such as peak height and area under the peak.The purpose of this study is to design a robust and adaptive classification method of MALDI-TOF mass spectra using deep learning and other techniques.This study built a model for MALDI-TOF mass spectra classification based on convolutional neural network,and determined the hyperparameters of the network model through a series of comparative experiments.Then this study used this model as a feature extractor combined with support vector machine,k-nearest neighbor,random forest and naive Bayes classifier for MALDI-TOF mass spectra classification.Secondly,inspired by the above work,a network model was built based on the denoising autoencoder as a feature extractor combined with other machine learning classifiers for mass spectra classification.Finally,this paper used k-fold cross-verification to evaluate the performance of the classification methods on a data set,which contained 8 major species and a total of 3355 MALDI-TOF mass spectra.The experimental results showed that the end-to-end classification method based on convolutional neural networks had the best effect.iIs accuracy rate,macro-averaged accuracy rate,macro-averaged recall rate,macro-averaged F1 score,weighted average recall rate,weighted average accuracy rate,weighted average F1 scores were all over 99%.Finally,based on the above works,this thesis designed and implemented a MALDI-TOF mass spectra identification software.The software uses a plug-in architecture to improve the scalability of software functions and facilitate later maintenance and development.