Mass classifier 338 -substructure

The applied scheme for the development of substructure classifiers can be applied to any substructures for which appropriate sets of mass spectra are available. Usually the same spectral features (variables) are used for different substructure classifiers. 

As was seen in the previous section, substructure classifiers are only one step in reducing the number of structures generated during CASE. Eigure 9.3 demonstrates that mass spectral match value and partitioning behavior reduced candidate numbers further, although many cases with more than 10 possible structures remained even within this small subset of 71 spectra. In terms of positive identification, even 10 possible structures is too many and would mean the purchase (or synthesis) of these reference compounds before a complete identification could be made. This is obviously... 

Before moving on any further, we first consider the effectiveness of the mass spectral classifiers in reducing the number of candidates for selection. Of the 29 compounds, 15 were already reduced to the final, small (< 35) number of candidates using 95% probability classifiers alone. No additional classifiers were necessary (or possible) to reduce this number further, nor did calculated properties separate these candidates further. This means over half these compounds required only the default substructure information to constrain structure generation adequately. In a further 4 cases, the reduction in candidates to their final number (again below 35) was achieved by using an additional classifier or other restrictions, and no further elimination occurred with calculated properties. One further compound experienced only a slight additional re-... 

It is not trivial to define or select substructures that should be considered for this purpose. For the STIRS system, considerable effort has gone into the search for substructures that can be successfully classified by the implemented spectral similarity search. The Mass-Lib system uses a predefined set of 180 binary molecular descriptors to characterize the similarity of structures. In most investigations a more or less arbitrary set of substructures, functional groups or more general structural properties (compound classes) has been considered. Self-adapting methods that automatically analyse the molecular structures in the hitlist (for instance by searching for frequent and large substructures) have not been used up to now in MS. 

The number of occurrences of a certain substructure in the hitlist is compared with the corresponding number for the library and a probability is derived for the presence of that substructure in the unknown. This classification method is a variant of the well-known -nearest neighbour classification . Each mass spectrum is considered as a point in a multidimensional space the neighbours nearest to the spectrum of the unknown correspond to the most similar reference spectra in library search. If the majority of k neighbours (k is typically between 1 and 10) contain a certain substructure then this substructure is predicted to be present in the unknown. A drawback of this approach is the high computational effort necessary for classifying an unknown because a full library search is required. The performance has been described by Stein (1995, see Further reading section) as sufficient to recommend it for routine use as a first step in structure elucidation . 

Table 4 Types of numerical spectral features a set of spectral features can be used to classify a mass spectrum as to whether a certain substructure or a more general structural property is present or absent in the molecule...

Simple yes/no classifiers without the possibility to refuse the classification answer are not adequate for the recognition of substructures from mass spectra. 

---