Cargando…
Multiple Compounds Recognition from The Tandem Mass Spectral Data Using Convolutional Neural Network
Mixtures analysis can provide more information than individual components. It is important to detect the different compounds in the real complex samples. However, mixtures are often disturbed by impurities and noise to influence the accuracy. Purification and denoising will cost a lot of algorithm t...
Autores principales: | , , , |
---|---|
Formato: | Online Artículo Texto |
Lenguaje: | English |
Publicado: |
MDPI
2019
|
Materias: | |
Acceso en línea: | https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6943725/ https://www.ncbi.nlm.nih.gov/pubmed/31847456 http://dx.doi.org/10.3390/molecules24244590 |
Sumario: | Mixtures analysis can provide more information than individual components. It is important to detect the different compounds in the real complex samples. However, mixtures are often disturbed by impurities and noise to influence the accuracy. Purification and denoising will cost a lot of algorithm time. In this paper, we propose a model based on convolutional neural network (CNN) which can analyze the chemical peak information in the tandem mass spectrometry (MS/MS) data. Compared with traditional analyzing methods, CNN can reduce steps in data preprocessing. This model can extract features of different compounds and classify multi-label mass spectral data. When dealing with MS data of mixtures based on the Human Metabolome Database (HMDB), the accuracy can reach at 98%. In 600 MS test data, 451 MS data were fully detected (true positive), 142 MS data were partially found (false positive), and 7 MS data were falsely predicted (true negative). In comparison, the number of true positive test data for support vector machine (SVM) with principal component analysis (PCA), deep neural network (DNN), long short-term memory (LSTM), and XGBoost respectively are 282, 293, 270, and 402; the number of false positive test data for four models are 318, 284, 198, and 168; the number of true negative test data for four models are 0, 23, 7, 132, and 30. Compared with the model proposed in other literature, the accuracy and model performance of CNN improved considerably by separating the different compounds independent MS/MS data through three-channel architecture input. By inputting MS data from different instruments, adding more offset MS data will make CNN models have stronger universality in the future. |
---|