Expert Algorithm for Substance Identification Using Mass Spectrometry: Application to the Identification of Cocaine on Different Instruments Using Binary Classification Models

Thisis the second of two manuscripts describing how general linearmodeling (GLM) of a selection of the most abundant normalized fragmention abundances of replicate mass spectra from one laboratory can beused in conjunction with binary classifiers to enable specific andselective identifications with reportable error rates of spectra fromother laboratories. Here, the proof-of-concept uses a training setof 128 replicate cocaine spectra from one crime laboratory as thebasis of GLM modeling. GLM models for the 20 most abundant fragmentsof cocaine were then applied to 175 additional test/validation cocainespectra collected in more than a dozen crime laboratories and 716known negative spectra, which included 10 spectra of three diastereomersof cocaine. Spectral similarity and dissimilarity between the measuredand predicted abundances were assessed using a variety of conventionalmeasures, including the mean absolute residual and NIST's spectralsimilarity score. For each spectral measure, GLM predictions werecompared to the traditional exemplar approach, which used the averageof the cocaine training set as the consensus spectrum for comparisons.In unsupervised models, EASI provided better than a 95% true positiverate for cocaine with a 0% false positive rate. A supervised binarylogistic regression model provided 100% accuracy and no errors usingEASI-predicted abundances of only four peaks at m/z 152, 198, 272, and 303. Regardless of the measureof spectral similarity, error rates for identifications using EASIwere superior to the traditional exemplar/consensus approach. As asupervised binary classifier, EASI was more reliable than using Mahalanobisdistances.

Automated summary

This study demonstrates the use of binary classifiers and GLM to enable specific and selective identifications of mass spectra from different laboratories, with lower error rates compared to traditional approaches.