Managing of Unassigned Mass Spectrometric Data by Neural Network for Cancer Phenotypes Classification

Abstract

Mass spectrometric profiling provides information on the protein and metabolic composition of biological samples. However, the weak efficiency of computational algorithms in correlating tandem spectra to molecular components (proteins and metabolites) dramatically limits the use of “omics” profiling for the classification of nosologies. The development of machine learning methods for the intelligent analysis of raw mass spectrometric (HPLC-MS/MS) measurements without involving the stages of preprocessing and data identification seems promising. In our study, we tested the application of neural networks of two types, a 1D residual convolutional neural network (CNN) and a 3D CNN, for the classification of three cancers by analyzing metabolomic-proteomic HPLC-MS/MS data. In this work, we showed that both neural networks could classify the phenotypes of gender-mixed oncology, kidney cancer, gender-specific oncology, ovarian cancer, and the phenotype of a healthy person by analyzing ‘omics’ data in ‘mgf’ data format. The created models effectively recognized oncopathologies with a model accuracy of 0.95. Information was obtained on the remoteness of the studied phenotypes. The closest in the experiment were ovarian cancer, kidney cancer, and prostate cancer/kidney cancer. In contrast, the healthy phenotype was the most distant from cancer phenotypes and ovarian and prostate cancers. The neural network makes it possible to not only classify the studied phenotypes, but also to determine their similarity (distance matrix), thus overcoming algorithmic barriers in identifying HPLC-MS/MS spectra. Neural networks are versatile and can be applied to standard experimental data formats obtained using different analytical platforms.

Analysis of Primary Liquid Chromatography Mass Spectrometry Data by Neural Networks for Plant Samples Classification

“…These heatmaps were cropped to a squared form (750 × 750) via rejection of the less-important signals in low (100–200 m / z ) and high (950–1300 m / z ) m / z regions. Compared to the approaches described in the literature [ 20 , 21 ], where a series of images was generated for each sample, representation of LC-MS data in the form of a single image seems promising because it saves spatial information and preserves the correlation between the adjoined signals in the sample. The squared images are ideal as input data for a convolutional NN with different losses and architectures.…”

“…In this case, supervised machine learning methods can be applied to provide a basis for sample discrimination. To our best knowledge, there has been no attempt regarding direct pattern recognition for plant material samples; however, this strategy has been employed to distinguish different types of cancer [ 20 , 21 ].…”

“…SVM can perform better if more preprocessing is applied to the data [ 16 ]. A 3D CNN was successfully applied in the analysis of raw (“unassigned”) HPLC-MS data for cancer phenotype classification [ 21 ]. Each file was represented as 98 images with dimensions of 512 × 512 pixels with color-coded intensities, thereby achieving a classification accuracy for several types of cancer of slightly above 90%.…”

“…The developed approach was applied to distinguish different plant parts from the Apiaceae family. Residual convolutional NN was chosen as the basic architecture for applied NN as they had already been used for mass-spectrometry data [ 21 , 27 ].…”

Deep learning facilitates multi-data type analysis and predictive biomarker discovery in cancer precision medicine

“…DL may assist in the in-depth study of cancer by enabling the integrative analysis of multi-omics datasets [48] , [49] , [50] , [51] . A generalized schema ( Fig.…”