Unsupervised Reconstruction of Analyte-Specific Mass Spectra Based on Time-Domain Morphology with a Modified Cross-Correlation Approach

Abstract

Concomitant species that appear at the same or very similar times in a mass-spectral analysis can clutter a spectrum because of the coexistence of many analyte-related ions (e.g., molecular ions, adducts, fragments). One method to extract ions stemming from the same origin is to exploit the chemical information encoded in the time domain, where the individual temporal appearances inside the complex structures of chronograms or chromatograms differ with respect to analytes. By grouping ions with very similar or identical time-domain structures, single-component mass spectra can be reconstructed, which are much easier to interpret and are library-searchable. While many other approaches address similar objectives through the Pearson’s correlation coefficient, we explore an alternative method based on a modified cross-correlation algorithm to compute a metric that describes the degree of similarity between features inside any two ion chronograms. Furthermore, an automatic workflow was devised to be capable of categorizing thousands of mass-spectral peaks into different groups within a few seconds. This approach was tested with direct mass-spectrometric analyses as well as with a simple, fast, and poorly resolved LC–MS analysis. Single-component mass spectra were extracted in both cases and were identified based on accurate mass and a mass-spectral library search.

Expert Algorithm for Substance Identification Using Mass Spectrometry: Statistical Foundations in Unimolecular Reaction Rate Theory

“…Sigman and Clark examined the two-dimensional cross-correlations in replicate spectra of high explosives, but the cross-correlations were examined as a function of a deliberate perturbation, i.e., differing collision energies, and the cross-correlations were not used to support compound identification, unlike the present work. In mass spectrometry applications, principal component analysis (PCA) has been used in two main ways: (1) to resolve or deconvolute mass spectra of mixtures, ,,, and (2) to relate a spectrum to other classes or structures within a database. ,− The latter approach has also been used in conjunction with discriminant analysis and binary classification algorithms to enable the classification of spectra to known identities. − Finally, machine learning and artificial intelligence methods have existed since the early 1970s, and they continue to be explored as methods to both identify known compounds in a library and to propose structures for compounds that are not in a library. ,,,,− Whereas the predictive power of sophisticated computational techniques is likely to continually advance, very few of the articles described so far tackle the difficult problem of discriminating between spectrally similar compounds collected on different instruments and without reference spectra from those instruments.…”

“…Sigman and Clark examined the two-dimensional cross-correlations in replicate spectra of high explosives, 89 but the cross-correlations were examined as a function of a deliberate perturbation, i.e., differing collision energies, and the cross-correlations were not used to support compound identification, unlike the present work. In mass spectrometry applications, principal component analysis (PCA) has been used in two main ways: (1) to resolve or deconvolute mass spectra of mixtures, 66,68,70,90 and (2) to relate a spectrum to other classes or structures within a database. 74,91−96 The latter approach has also been used in conjunction with discriminant analysis and binary classification algorithms to enable the classification of spectra to known identities.…”

“…One aspect of spectral comparison algorithms that is often overlooked is the assumption, and oftentimes mathematical requirement, that any variance in the relative abundance of peaks within a replicate spectrum is randomly or independently variable at each m / z value. , Such a mathematical requirement has been assumed since the first use of computational approaches to background-subtraction or spectral deconvolution into discrete component spectra, ,, whether using simultaneous linear equations or matrix theory. , By default, deconvolution algorithms explicitly assume unit correlation among the absolute signals of fragments as a function of time, or scan number, and they implicitly assume that any unexplained variance in a given scan at a specific m / z value is random. ,,,,,,− Furthermore, to have statistical validity, most measures of spectral similarity and dissimilarity between questioned and reference spectra also require independent variance, i.e., no correlation, in the relative abundance at each m / z value within replicate spectra. ,, As an example of this reliance, a recent and extremely effective approach to spectral comparisons uses combined unequal variance t -tests at each m / z value to compare questioned and known spectra. Combining the results of independent t -tests explicitly requires independent variability of each t -test to enable the computation of random match probabilities. − However, as indicated elsewhere, and as we will show below, replicate spectra still contain strong correlations in the normalized abundances of peaks, so the different m / z values are not independently variable.…”

Cumulative Neutral Loss Model for Fragment Deconvolution in Electrospray Ionization High-Resolution Mass Spectrometry Data

“…Currently, the fragment deconvolution approaches (in DIA experiments) that are suitable for NTA focus heavily on the information presented in the time domain. ,,, One of these approaches is the apex retention time matching of the precursor features with the potential fragment signals. ,,,,, In addition to this, some algorithms also use peak shape assessment via correlation analysis to group fragments. ,− ,, More decomposition-based algorithms (e.g., MCR and/or PARAFAC) are available to perform signal deconvolution. ,, These methods are generally less suitable for NTA due to the requirement of multiple samples, which often means that an analyte needs to be present in more than one sample and at different concentrations for the method to work. However, since it is unknown what is present in the samples to begin with, ensuring that the presence of this compound across multiple samples becomes impossible, ,,, even more so if different concentrations of the analyte are required to resolve the mass spectrum .…”

“…4,11,13,16,17,25 In addition to this, some algorithms also use peak shape assessment via correlation analysis to group fragments. 4,[16][17][18][19]26,27 More decomposition-based algorithms (e.g., MCR and/or PARAFAC) are available to perform signal deconvolution. 24,26,28 These methods are generally less suitable for NTA due to the requirement of multiple samples, which often means that an analyte needs to be present in more than one sample and at different concentrations for the method to work.…”