ИСТИНА

HPLC-MS method is most often used for the potential chemotaxonomic markers search in plants extracts due to its informativeness, sensitivity and lack of preliminary derivatization. In our work we obtained mass-chromatogram of samples from 19 species of the Apiaceae family. In the first part various unsupervised algorithms for multivariate statistical analysis were used to process the obtained data such as dimensionality reduction methods (PCA, ICA, NMF, PARAFAC) and unsupervised feature selection (UFS). Obtained clustering results (dendrograms) were compared with biological taxonomic tree. Also, 23 potential chemotaxonomic markers for plants from the Apiaceae family have been identified. These markers were mostly related to the coumarin group. Their preliminary identification was carried out using HPLC-HRMS (MS and MS2 spectra)1. However, these biomarkers don’t identify groups of plants by their genus. In the second part of the work supervised machine learning methods were applied to the extended with other plants’ parts dataset for classification and biomarkers’ search. Basic approaches already widely described in literature2 like PLS-DA and SVM were applied to unfolded dataset. After revealing class-specific features it was found out that they are not unique for each class. Further it was decided to try to apply neural networks because they are able to handle LC-MS data without vectorization and are known for huge flexibility and high level of accuracy. Because of the specific constraints in the dataset like class imbalance, low number of samples and expectation of specific features Siamese network with triples loss function was used. Moreover, dataset was extended using augmentation techniques with respect to LC-MS data specific. Results obtained by this method were compared with mentioned basic machine learning techniques. HPLC-MS-TOF analysis for biomarker identification was performed using the equipment of the demo laboratory of Bruker Ltd., Moscow, Russia