Research On Risk Related Markers In The Thermal Treatment Process For Liquid Milk Using Mass Spectrometry And Machine Learning

This article explores the effects of different heat treatment conditions on peptides and flavor compounds in milk,and identifies potential heat-sensitive biomarkers.Additionally,a classification model for different heat-treated milk was established using machine learning algorithms.The specific research results are as follows:(1)Research on the association of multiple peptides in milk heat treatment based on MALDI-TOF-MS combined with machine learning.Organic precipitation protein pre-treatment was used to extract peptides,and MALDI-TOF-MS was used to obtain peptide mass spectrometry data.By comparing the peptide mass spectra at different temperatures,it was found that the concentration range of the mass spectrometry signal was between m/z 900-2500,and the number of peptides increased with increasing temperature.LASSO,RFECV,and PLSDA were used for feature extraction of peptide mass spectrometry data,and 30 peptide segments related to temperature were selected.Six heat-sensitive peptide segments were identified,namely m/z 923,2107,2145,2331,2910,and 3706.Four machine learning algorithms,SVM-L,RF,LDA,and PDA,were found to have the best performance,with an accuracy of 97 %,96 %,.96 %,and 96 %,respectively.In blind sample prediction,18 out of 20 samples were correctly predicted.(2)Research on the association of flavor substances in milk heat treatment based on GC-MS combined with WGAN-GPSolid-phase microextraction(SPME)was used to adsorb flavor compounds in milk,and GC-MS was used to obtain total ion chromatograms(TIC)for flavor compound identification.A total of 48 milk flavor compounds were identified,including 6 alcohols,5 lipids,2 phenols,13 acids,3 alkanes,8 ketones,5 aldehydes,and 6 other compounds.The results showed that the number of flavor compounds in milk increased with increasing heat treatment temperature.WGAN-GP was used for data augmentation,and the results showed that the amplified samples could simulate the variable distribution of the original data,but there were still some variables that deviated from the original variable distribution,which increased the diversity of the sample data.Seven machine learning models were established,and the algorithm performance was compared before and after data augmentation.The results showed that six algorithms could improve the algorithm model performance after data augmentation.The WGAN-GP combined with the deep neural networks algorithm model had the best performance,with an accuracy of 96 %.In the prediction of 15 blind samples,13 samples were correctly identified.