MS classifiers

A library of mass spectra with the corresponding compounds represented as molecular graphs is needed for the development of MS classifiers. To construct an MS classifier 

mass spectra are mapped onto real numbers by MS descriptors D = 

Descriptor values necessary or helpful for training the predicting function have to be transformed 

In a previous approach [335], LDA, KNN, ANN tuid soft independent modeling of class analogy (SIMC A) were tested and compared, and ANN and LDA proved to be preferable. In the following we shall calculate classifiers via CART and LDA, and then compare them with those obtained by SVM and ANN. 

A base set of 86,052 spectrum-structure pairs from the NIST MS library (Subsection 8.3.4) was scanned for several structural properties contained in Appendix B. For a total of 77 properties there were at least 300 structures with euid at leeist another 300 structures without the given property. Disjoint leeuning and test sets were selected randomly, 150 with and 150 without the property. 

---

Figure 8.4 shows our workflow for structure elucidation via MS, following the plan, generate, test strategy used in DENDRAL (Section 8.2). The focus is on determination of the molecular formula and structure. For interpretation we use MS classifiers, which provide information on both element composition and structure (see Appendix B). We use classifiers described by K. Varmuza and W. Werther [324, 333, 335] and develop new classifiers based on different classification methods (Subsection 8.5.2) and new structural properties (Subsection 8.5.3). 

A database of elucidated spectra is indispensable for both quality assessment of ranking functions and for calculation of MS classifiers. Here, we use spectra and structures from the NIST MS library [224]. This 1998 version of NIST contains 107,888 spectra of 107,812 structures. Spectra and structures are two separate files, linked by numerical identifiers. 

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. 

Figure 8.25 shows the procedure to calculate and apply an MS classifier. Note that this principle is not restricted to mass spectrometry. For example, in [232] the construction of IR classifiers is described using the same scheme. 

Table 8.12. Misclassification rates of MS classifiers (CT) for 77 structural properties.

Among the 1790 substructures there are 301 for which such sets can be obtained. For each of these 301 substructures we calculate MS classifiers. For descriptor selection we use, as described above, the 50 -fold stepwise procedure within MLR. By so doing we obtain 13 MS descriptors relevant for modeling for each substructure. For classification CT, LDA, ANN with one, two, or three HN, and SVM with linear, radial, polynomial (degree = 2), and sigmoid kernel are used. 

We use the MS classifiers from MSclass [324] to extract structural properties. These classifiers were obtained via classification by regression (Subsection 6.1.1). In training the classifiers for the values of the discriminant function, percent quantiles were calculated. As a result, it is possible to give an estimated precision for each prediction made such that classifiers of low precision can be excluded from further consideration. Unfortunately, by doing so the number of structural properties for generation of molecular and structural formulas will be reduced. 

The 160 MS classifiers fi om MSclass return 6 positive and 42 negative answers of at least 95% reliability. The positive results are... 

E. Schymanski, C. Meinert, M. Meringer, and W. Brack. The use of MS classifiers and structure generation to assist in the identification of unknowns in effect-directed analysis. Anal. Chim. Acta, 615 136-147, 2008. 

---