Prediction of molecular substructure using mass spectral data based on metric learning

Abstract

In this paper, some metric learning algorithms are used to predict the molecular substructure from mass spectral features. Among them are Discriminative Component Analysis (DCA), Large Margin NN Classifier (LMNN), Information-Theoretic Metric Learning (ITML), Principal Component Analysis (PCA), Multidimensional Scaling (MDS) and Isometric Mapping (ISOMAP). The experimental results show metric learning algorithms achieved better prediction performance than the algorithms based on Elucidation distance. Contrasting to other metric learning algorithms, LMNN is the best one in eleven substructure prediction. © 2014 Springer International Publishing Switzerland.