Classification of mass spectra

For the statistical considerations in the previous section we restricted ourselves to structure spaces of no more than 10,000 constitutions for a given molecular formula. In practical applications, however, such cases will be the exception rather than the rule (see Appendix D). Thus, it should be possible to determine structural properties (SP) of the analyte prior to structure generation, so that these can be used to restrict the number of generated structures. MS classifiers provide an opportunity to extract information on present or absent SP from mass spectra. 

A prerequisite for construction of a spectrum classifier is a database of elucidated spectra containing a sufficient number of structures with and without property SP. The presence of SP is the target variable for a statistical learning program for classification. 

In the case of mass spectra, it would be tempting to use peak intensities as predictors. However, intensities themselves are not linked strongly to structural properties. Instead, MS descriptors are more appropriate to model MS-structure relationships. Classification yields a prediction for the spectrum of an unknown that determines whether SP is to be considered as prescribed or forbidden in the further course of structure elucidation. 

---

W. Werther, H. Lohninger, F. Stand and K. Vermuza, Classification of mass spectra. A comparison of yes/no classification methods for the recognition of simple structural properties. Chemom. Intell. Lab. Syst., 22 (1994) 63-67. 

Stonham TJ, Aleksander I, Camp M, Pike WT, Shaw MA (1975) Classification of mass spectra using adaptive digital learning networks. Anal Chem 47 1817... 

The strictly mathematical basis of the quasi-equilibrium rate theory is absent from other, qualitative theories of mass spectrometric fragmentation which are based mainly on empirical rules and are discussed more fully in Section VI on the classification of mass spectra. One empirical classification, the charge-localization treatment, has achieved an aura of theoretical respectability through the use of curved arrows and analogies with the formalism of resonance theory in solution chemistry (Budzikiewicz et al., 1967a). The bonds in a molecular ion vibrate with excess energy but since an electron has been removed on ionization, the vibration frequencies are not the same as in the original intact molecule. 

The first analytical application of a pattern recognition method dates back to 1969 when classification of mass spectra with respect to certain molecular mass classes was tried with the LLM. The basis for classification with the LLM is a discriminant function that divides the -dimensional space into category regions that can be further used to predict the category membership of a test sample. 

In the recent past SVM have been increasingly used to solve problems in computational chemistry. In a comparison of SVM and ANN for classification of pharmaceutically inactive or active compounds, SVM consistently yielded smaller classification errors [41]. For the classification of mass spectra (see Subsection 8.5.2), SVM with a radial kernel proved to be the best predicting functions. 

A plethora of other descriptors was tested during the development of the software MSclass [324] for the classification of mass spectra. MSclass finally contained 160 classifiers using 32 descriptors for 431 combinations of parameters in total. 

Along with CT and LM, we want to test ANN and SVM for classification of mass spectra. The selection of descriptors for constructing the tree is performed with the CART algorithm. Due to the intrinsic rapidity of MLR, many descriptor subsets are tested within a reasonable time and good ones are selected. For ANN and SVM such opportunities are not available. Nevertheless, in order to use these methods, we calculate ANN and SVM for the descriptor subsets determined by CT and MLR. Thus, we test ANNs of one, two, and three hidden neurons, and SVMs with linear, radial, polynomial (degree = 2), and sigmoid kernel along with CT (mindev = 0.04) and LDA for the 77 structural properties. 

K. Varmuza et al. MSclass. Software for Chemical-Structure-Structure-Oriented Classification of Mass Spectra. Classifier Guide. Applied ChemoMetrics, Technische Universitat Wien, 1996. 

The most important results and facts, in particular the extensions of the MOLG EN class library Reaction-based generation of structures, QSPR studies using different kinds of molecular descriptors, various methods for prediction, ranking and classification of mass spectra, relations between spectra and properties and CASE using electron impact (El) mass spectrometry. 

A mass spectral classifier is a part of a computer program that uses the peak list of a low resolution mass spectrum as input and produces information about the chemical structure as output. For such a classification procedure a number of methods are available in multivariate statistics. Many of them have already been applied to various problems in chemistry classification of mass spectra (with the aim of recognizing chemical compound classes) was one of the pioneering works in chemometrics. 

---