Evaluation of automatically generated substructure identification rules from tandem mass spectra

Abstract

Substructure identification rules for phenothiazine and barbiturate substructures were generated by using a new version of the Method for Analyzing Patterns in Spectra (MAPS) software. This software uses tandem mass spectra and known substructure content of reference compounds to provide "feature-combination" rules. A feature-combination is a series of tandem mass spectral features which are completely unique to compounds containing a specified substructure. The current reference databases contain over 11,000 daughter spectra of 100 compounds acquired at two different collision gas pressures (i.e., single- and multiple-collision conditions). The results of rule evaluation procedures are presented and include a comparison of the spectral features developed in rule generation to those identified in documented fragmentation pathways of the indicated substructure. Two potential sources of error due to spectral feature and substructure "cross-correlation" were identified. If errors occur, they can be detected by calculating cross-correlation coefficients and edited from the rules. A beneficial cross-correlation involving feature-combinations was also discovered. The rules obtained by using single- and multiple-collision data were further evaluated by applying them to tandem mass spectra of 20 test compounds (compounds not in the reference database). The results of these evaluations give a good indication of the utility of the rules for use in an automated structure elucidation system for tandem mass spectrometry data.

Generation of substructure identification rules using feature-combinations from tandem mass spectra

“…Although the excluded feature-combination would not have improved the identification of the substructure in any of the reference compounds, it is possible that it might enhance recall with an unknown compound. This situation was not observed for the phenothiazine and barbiturate rules that were applied to 20 test compounds [10]. These rules had 100% recall with respect to test compounds so no loss of information was observed using the HITS and Cc parameters.…”

“…The incorporation of feature-combinations in MAPS, on the other hand, has provided a rule for phenothiazine that has both a reliability and a recall of 100% with respect to the reference database. The reliability and recall of this rule was also tested against 20 test compounds and 100% reliability and recall were obtained [10]. Although these figures of merit do not mean that a MAPS rule will predict the presence of the phenothiazine substructure with absolute certainty in a true unknown, they do show that the new MAPS software is utilizing the M5/MS data more effectively than previous versions of the software.…”

“…It is expected, however, that the reliability and recall estimates will more accurately portray the effec- tiveness of the MAPS rules in analyzing true unknowns when a larger tandem mass spectral database is available to generate the rules. A detailed evaluation of the phenothiazine feature-combination rule is provided in the following article [10].…”

“…They can be used to identify the most important features or combinations of features that are associated with particular compound classes (i.e., substructures). In some cases, individual feature-combinations can lead to the identification of tandem mass spectral features that are indicative of combinations of substructures [10]. The MAPS rules can also be applied to the MS/MS data sets of unknown compounds to identify the presence of substructures within the molecular structure of the unknowns.…”

“…Two evaluations of the MAPS rules have been performed to investigate the performance of the new MAPS software. These include a comparison of the spectral features selected for the phenothiazine and barbiturate substructures using the Ui/Ci criteria and application of the resulting feature-combination rules to a series of test compounds, as reported in the next article [10]. One of the major conclusions of these evaluations was that optimal rule reliability was achieved using a MINF value > 2.…”