Article

pubs.acs.org/ac

Multivariate Curve Resolution-Alternating Least Squares Analysis of

High-Resolution Liquid Chromatography−Mass Spectrometry Data
Melanie M. Sinanian,† Daniel W. Cook,† Sarah C. Rutan,*,† and Dayanjan S. Wijesinghe‡
†
Department of Chemistry, Virginia Commonwealth University, Richmond, Virginia 23284, United States
‡
Department of Pharmacotherapy and Outcomes Science, Virginia Commonwealth University, Richmond, Virginia 23284, United
States
*
S Supporting Information
See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.

ABSTRACT: Methods such as liquid chromatography coupled

Downloaded via INDIAN INST OF TECH DELHI- IIT on July 14, 2021 at 20:33:02 (UTC).

with high-resolution mass spectrometry (LC-HRMS) are crucial

for differentiating compounds with highly similar masses. This is a
necessity when analyzing highly complex samples; however, the
size of high-resolution LC-HRMS data sets can cause difficulties
when applying advanced data analysis techniques. In this work,
LC-HRMS analyses of known amphetamine samples and
unknown bacterial lipid samples were carried out, and multivariate
curve resolution-alternating least squares (MCR-ALS) was
applied to the data to obtain mathematical separation of
overlapped analyte signals. In order to minimize computational
strain, a novel strategy was developed which minimizes the
number of irrelevant masses analyzed at full resolution. To do
this, data were first binned to unit mass resolution, and MCR-ALS was performed. This provided mathematical components for
each analyte present plus background components. In the resolved spectral profiles of analyte components, masses above a preset
intensity threshold were extracted, discarding all other masses, and expanded to successively higher levels of resolution, applying
MCR-ALS at each level. These steps were repeated until 0.001 amu resolution was achieved, as dictated by the resolution of the
instrumentin this case, a time-of-flight mass spectrometer. This strategy allowed for the accurate recovery of all known
amphetamine compounds and select bacterial lipid extracts while minimizing the size of the data, therefore minimizing
computational analysis time and data storage requirements. This relatively simple strategy enables the effective coupling of LC-
HRMS with MCR-ALS.

T he combination of liquid chromatography with high-

resolution mass spectroscopy (LC-HRMS) offers a
powerful tool for the resolution of peaks in two modes
temporal and spectral. This offers the potential for qualitative
and quantitative analysis of highly complex mixtures. Despite
the temporal and spectral resolving power of LC-HRMS,
overlapping and interfering signals are often present in both
domains, especially in cases where the sample is highly
complex. Another complication is the inclusion of irrelevant
signals in the data. Because of the sensitivity of these
instruments, much of the data collected in LC-HRMS
experiments correspond to noise and background rather than
true analyte signals. Because of this, it is important to efficiently
differentiate analyte signals from irrelevant signals, a process
known as peak detection. Many different algorithms have been
implemented in many software packages that approach peak
detection differently. Many of these rely on the Gaussian-like
peak shapes in both the chromatographic and spectral modes,
including direct fitting of Gaussian or exponentially modified
Gaussian peak models to the extracted ion chromatograms.1,2
Another approach is the use of derivatives to remove noise and
linear background signals while defining peak start and end
points.3−6 While these approaches are widely used, they can
often struggle to detect low-level analytes, due to required
intensity thresholds. Furthermore, while such approaches can
often resolve slightly overlapped signals they are often unable
to resolve highly overlapped signals.
An alternative to peak detection is the use of multiway
chemometric methods to resolve individual analyte signals from
both background and other analyte signals. The data resulting
from LC-HRMS experiments are considered to be “second-
order” data, meaning that the data can be represented by a
matrix with time in one mode and mass-to-charge (m/z) in the
second mode. Chemometric data analysis methods employ
mathematical techniques that are able to handle second-order
and higher dimensional data, allowing the entire data set to be
analyzed in a single analysis rather than analyzing a single m/z
channel at a time (e.g., extracted ion chromatograms). One
method that has been developed to resolve chromatographic
data is multivariate curve resolution-alternating least squares
Received: August 10, 2016
Accepted: October 18, 2016
Published: October 18, 2016

© 2016 American Chemical Society 11092 DOI: 10.1021/acs.analchem.6b03116

Anal. Chem. 2016, 88, 11092−11099
Analytical Chemistry
(MCR-ALS).7−9 MCR-ALS has been widely used for second-
order data arising from LC with ultraviolet−visible detection
(UV−vis),10−14 LC with fluorescence spectroscopy,12,15,16 and
with low-resolution mass spectrometry.17,18 MCR-ALS has also
recently been utilized for LC-HRMS analyses, including
metabolomics;19 however, the advantages that the high-
resolution data provide have not been fully realized. In almost
all cases, the LC-HRMS data analyzed by MCR-ALS is
subjected to binning, a process of grouping the mass intensities
into bins within a specific range.3 This is done to reduce the
size of the LC-HRMS data, which is very large due to the
number of masses in the data set. For example, a mass spectrum
with a range from 50 to 1000 amu at intervals of 0.001 amu
contains a possible 9 × 106 mass-to-charge (m/z) values.
Binning to intervals of 0.1 amu, for example, reduces the
number of possible m/z values to 9 × 103.
Even though unmanageable in its raw form, the information
contained in HRMS data is often necessary to differentiate
between compounds with very similar masses. In a lipidomic
study of placental cells by Gorrochategui et al., MCR-ALS was
performed on binned LC-HRMS data. After the data was
resolved at unit mass resolution, the authors examined the raw,
high-resolution data at masses found to correlate to potential
biomarkers. From the raw data, masses at 0.0001 amu precision
were assigned.20 This approach can be problematic if
chromatographically overlapped species contain spectral peaks
which share the same mass at lower precision because MCR-
ALS would be unable to resolve these peaks in low precision
data.13,21,22 An alternative approach to binning is the use of
wavelet transforms; however, when MCR-ALS is to be used,
the Haar wavelet must be selected to retain non-negativity in
the data. The effect of the Haar wavelet transformation is a
pairwise averaging effect which is practically equivalent to the
binning process, with a loss of resolution accompanying the
compression. Recently, Tauler et al. have published a new
protocol outlining a different approach to LC-HRMS data
analysis using MCR-ALS.23 Their approach defines “regions of
interest” (ROI) prior to MCR-ALS analysis, allowing for data
compression to take place.24 These regions of interest are
chosen on the basis of several parameters including a signal-to-
noise threshold, which if set incorrectly may lead to the
exclusion of compounds at low intensities, particularly if a low-
level analyte is present in the vicinity of a much higher
concentration analyte.
The work described in the current paper presents a new
strategy which can analyze LC-HRMS data by finding relevant
regions in the binned mass spectra using MCR-ALS and
discarding all other masses. These regions are used for a second
round of MCR-ALS at a 0.1 amu bin level. This process is
repeated until data at 0.001 amu precision are analyzed. This
allows for the resolution of compounds which overlap in the
chromatographic mode and share masses at even 0.01 amu
precision, while greatly reducing the size of the data. In contrast
to the ROI approach, the current approach does not make use
of any thresholds prior to MCR-ALS analysis, allowing for low-
level analytes to be captured without risk of mistakenly
eliminating them due to incorrect thresholding.
■ THEORY
Multivariate Curve Resolution-Alternating Least
Squares (MCR-ALS). MCR-ALS is an iterative optimization
method that mathematically resolves signals arising from
chemical species and background without needing prior
11093
Article
information about the spectra or concentrations of the
compounds present in the sample.19,8,22 MCR-ALS can be
viewed as a multicomponent Beer’s law relationship given as
the equation below.
X = CST (1)
In this relationship, X is the raw second-order data resulting
from an LC-MS run, C is a matrix consisting of vectors
representing the pure chromatographic profiles, and ST is the
corresponding matrix of pure mass spectra.3,8 An initial guess
for either the spectral or chromatographic profiles allows for the
solution of eq 1 for either C or ST using alternating least-
squares algorithms.4,25 Most often, spectral initial guesses are
used, which can be obtained through methods such as
SIMPLISMA (ACD/Laboratories, Toronto, Canada)26 or
iterative orthogonal projection analysis (IOPA),27−29 both of
which aim to extract the most dissimilar spectra from the raw
data. In this work, IOPA was used. Briefly, IOPA extracts a
certain number of spectra, defined by the user, from the raw
data that are the most orthogonal from one another.
Constraints. The defining step in MCR-ALS is the
application of constraints to drive the solution toward the
correct, chemically relevant answer for the pure, resolved
components. Commonly employed constraints include non-
negativity, selectivity, unimodality (one maximum per
component), closure (mass balance), smoothness,30 compo-
nent correspondence,31 area correlation,32 and hard modeling
constraints.8,9 Of the constraints listed above, nonnegativity and
selectivity were applied in this procedure. Nonnegativity
ensures that the chromatographic and mass spectral intensities
do not go below zero, as the negative values are not physically
reasonable. Selectivity notifies the algorithm of prior known
values, such as regions of time known to contain no
compounds or regions of the mass spectra known to contain
no signals.25,33 These constraints were applied only to those
components believed to correspond to true chemical species.
Additional constraints beyond these provided no improvement
to the resolution results.
Extended MCR-ALS. As shown in eq 1, MCR-ALS operates
on second-order data (i.e., a matrix); however, MCR-ALS is
also capable of analyzing multiway (e.g., multiple samples,
multidimensional chromatography,34,35 etc.) and multiset data
(e.g., data fusion36). Making use of this higher-order data can
greatly reduce the amount of rotational ambiguity often
associated with MCR-ALS.37 For example, we often want to
include multiple samples, creating a multiway array (i.e., a
cube) of data. In order to analyze all samples simultaneously,
augmentation of the third-order array into a second-order array
is necessary. This process is illustrated in Figure 1. Essentially,
every sample is concatenated along the time mode to create a
single augmented time mode containing all samples, while
conserving the spectral mode. When this augmented data
Figure 1. Graphical representation of data rearrangement process for
reshaping a third-order data array to a second-order data array.
DOI: 10.1021/acs.analchem.6b03116
Anal. Chem. 2016, 88, 11092−11099
Analytical Chemistry
matrix is analyzed with MCR-ALS, it is called extended MCR-
ALS.25
■ STRATEGY
The overall strategy proposed for the analysis of LC-HRMS
data is outlined in Figure 2. The approach includes selection of
Figure 2. Summary of strategy for analysis of LC-HRMS data by
MCR-ALS.
retention time windows, binning the spectra to unit mass,
MCR-ALS analysis, followed by selection of relevant masses
from this analysis and examining these at 10-fold higher mass
precision. MCR-ALS is then carried out on these data, and the
procedure is continued until analysis at the resolution precision
of the instrument is achieved.
1. Selection of Retention Time Windows. The first step
in most MCR-ALS analyses, including the present one, is to
select windows along the time mode in order to minimize the
complexity of the data submitted to the MCR-ALS analysis
(i.e., number of compounds, data size). This is shown in Figure
3. Time windows are chosen such that they are the same across
all samples so that they ideally contain the same analyte peaks
in each sample. The following steps were applied to each
window individually. If peak windows cannot be chosen in
which no peaks are split across the window edge, windows can
be chosen in which the regions overlap to ensure that each peak
is completely contained in at least one window. This procedure
of segmentation is inevitably used in MCR-ALS applications in
chromatography because of the localization of the component
information within narrow time windows.
2. Binning Mass Spectra to Unit Mass. As described in
the previous section, analyzing the entire mass range in LC-
HRMS data requires binning the data to decrease the number
of data points and therefore reduces the size of the data needed
to be analyzed. In this strategy, the data is initially binned to
unit mass resolution. The binning procedure collects masses
around a certain value and sums the intensities and assigns
them to that value. Because small molecules tend to have a
mass defect ∼0.1 amu, an asymmetric binning process, was
utilized for the initial, unit mass binning procedure. For
example, at unit mass, the intensity at all masses between
149.6000 and 150.5999 amu were summed and assigned to 150
amu. This process is often performed prior to MCR-ALS
analysis of LC-HRMS data, eliminating the advantages of high-
resolution data. The current strategy makes use of several
binning steps at sequentially higher resolutions.
3. MCR-ALS Analysis. The binned data is then augmented
to include all samples, as described in the Theory section above,
11094
Article
Figure 3. Total ion chromatogram (TIC) (A) and contour plot (B)
for an amphetamine chromatogram depict the partition of data in the
time mode into the analyzed windows, differentiated by numbered
boxes. The two peaks between 150 and 190 s were diastereomers and
therefore are unable to be resolved via MCR-ALS because they have
identical spectra.
and an initial guess is obtained using IOPA to initiate MCR-
ALS analysis. For the initial unit mass resolution analysis, the
number of components is chosen on the basis of the expected
number of compounds plus one or two background
components. Scree plots38 are often used as a starting point
for determination of the number of components in MCR-ALS.
For chromatographic data, however, scree plots are often
difficult to interpret due to a lack of distinct “shoulders” or
breaks in the graph which are used to estimate the number of
components.28 In the case of untargeted analyses, several
numbers of components should be tried, and the most
reasonable number of components (based on peaks split
between components and overall realistic looking background/
analyte profiles) should be chosen. Because the initial round of
MCR-ALS should eliminate the majority of background ions,
the number of components at analyses at subsequent bin levels
should be less than at unit mass resolution. In many cases, the
number of components at these subsequent levels will equal the
number of compounds present, without any background
components, because of the elimination of the background
masses in the initial round of MCR-ALS. The chromatographic
profiles and the mass spectral profiles of each component are
reviewed, and those which appeared to correspond to true
chemical components are chosen for submission to the next
step. Any components which may be difficult to assign as a
chemical component versus a background should be included
in the next step to prevent falsely excluding true analyte peaks.
4. Expansion of Relevant Masses to Next m/z Level
and Extraction from Raw Data. To prepare for higher-
resolution analysis in subsequent steps, masses with intensities
DOI: 10.1021/acs.analchem.6b03116
Anal. Chem. 2016, 88, 11092−11099
Analytical Chemistry Article

greater than a set threshold percentage of the spectral peak
intensity within each analyte’s resolved spectral prof ile are chosen
as significant. In this work, a threshold of 5% of the maximum
was used. It is important to note that the threshold is applied to
each component individually, negating any possibility of
excluding low-intensity compounds. Compounds at low
intensities should resolve into their own component and their
spectra will be normalized to the maximum intensity, and thus,
masses significant to those compounds should dominate their
corresponding spectral profile regardless of their intensity in the
raw, unresolved data. The significant masses are then expanded
to the next bin level (e.g., from 0.1 to 0.01 amu) and signals
corresponding to those masses were then extracted from the
raw data and subsequently binned to the new bin size similar to
the binning procedure in step 1. For example, if a component
contains a significant mass at 163 amu, for the next step of the
analysis, the masses selected for binning cover a range from
162.6 to 163.5 at 0.1 amu intervals. From the raw data, signals
at masses 162.5500−162.6499 amu are binned to 162.6 amu
and 162.6500−162.7499 are binned to 162.7 and so on. This
process is portrayed graphically in Figure 4. Only data at these
computer while the former was used for data translation and
program development. The Bioinformatics Toolbox by Math-
works was used for importing data into MATLAB.
Data Collection. Two data sets were analyzed with the
strategy described above to demonstrate its applicability to both
targeted analyses and untargeted, discovery type analyses. Both
sample sets were analyzed with a Shimadzu LC system (Nexera
series, Kyoto, Japan) coupled to an AB Sciex TripleTOF 5600
mass spectrometer (Concord, Ontario, Canada). The chroma-
tographic conditions for each sample set are listed in their
respective sections below. The data were collected in profile
mode, rather than being centroided. Data were converted from
AB Sciex .wiff files to mzXML files using msConvert, which is
contained in the ProteoWizard suite.9 These data were then
imported into MATLAB using the Bioinformatics toolbox
mzxmlread function, and peak lists were extracted. These were
then ready to be processed using the in-house binning program.
Amphetamine Samples. Amphetamine standards were
purchased from Grace Discovery Services (Columbia, Mary-
land). The names, abbreviations, and structures of the
amphetamines used are listed in Figure 5. The compounds
were divided into three groups. For each group, four calibration
mixtures and two test mixtures were created. The concen-
trations and further sample information are given in Supporting
Information Table S1. The data from all three groups were
analyzed as a single data set. This analysis represents an ideal

Figure 4. From the binned data, MCR-ALS extracts relevant masses.

These masses are then expanded to a higher resolution level and
MCR-ALS again extracts relevant masses at this resolution level. The
blue and red boxes represent masses in the true chemical components,
and the gray boxes represent background or masses with insignificant
intensities.

extracted and expanded masses (above the 5% threshold in this

work) are analyzed in the next round, keeping the size of the
overall data set to a level that is reasonable for analysis. This
process eliminates masses that do not correspond to chemical
compounds and therefore are irrelevant for analysis.
5. Repeat Steps 3 and 4 until Desired Final m/z Level
Is Reached. MCR-ALS analysis (with IOPA initial guesses
each time) and extraction of masses as described in steps 3 and
4 are repeated, stepping down to a higher resolution level after
each iteration. Once the data are expanded to the desired final
resolution level as dictated by the instrumental data (e.g., 0.001
amu), the spectral peaks can then be used for compound
identification, while the resolved chromatograms can be used
for quantitative analysis and/or pattern recognition.
The application of this general strategy is demonstrated in
the following sections. Examples of both a targeted analysis of
amphetamines and an untargeted analysis of bacterial lipid
samples are shown.

■ EXPERIMENTAL SECTION
Software. All programs were written in-house using
MATLAB (Mathworks, Inc., Natick, MA) version R2013a on
a Dell Precision T3600 with an Intel Xeon E5−1620 CPU at
3.60 GHz and 32.0 GB of RAM and version R2015b on a Dell
Optiplex 9020 with an Intel Core i7−4790 CPU at 3.60 Hz and
32.0 GB of RAM. Both systems were running Windows 7 Figure 5. Structures and abbreviations of amphetamines contained in
Enterprise. Most of the calculations were run on the latter the amphetamine standard solutions analyzed.

11095 DOI: 10.1021/acs.analchem.6b03116

Anal. Chem. 2016, 88, 11092−11099
Analytical Chemistry
analytical experiment and is used to demonstrate the feasibility
of our strategy.
For the amphetamine analysis, an Accucore C18 column (2.1
× 100 mm, 2.6 μm; Thermo Scientific, Waltham, MA) was
used. Acetonitrile was used as mobile phase B, and 10 mM
formic acid was used as mobile phase A. Gradient elution was
used starting at 2.5% B, followed by an increase to 35% B over
10 min.
Bacterial Lipid Analysis. To demonstrate the utility of our
strategy to complex analyses, five replicates from three different
strains of bacteria were analyzed. To prepare for analysis, the
samples were freeze−thawed three times in 200 μL of
phosphate-buffered saline followed by probe sonication.
Then, 1 mL of methanol was added followed by bath
sonication. Another round of sonication was performed with
0.5 mL of chloroform. After a 2 h incubation at 48 °C, 1 mL of
chloroform and 3 mL of water were added followed by
vortexing and centrifugation. The organic layer was extracted. A
second extraction was performed with an additional 2 mL of
chloroform. The organic extract was dried via vacuum
centrifugation and resuspended in 100 μL of methanol for
analysis.
These samples were analyzed using an Acuity HSS T3 C18
column (2.1 × 150 mm, 1.8 μm; Waters, Milford, MA) at 55
°C. Gradient elution was used with mobile phase A consisting
of 60:40 water:methanol with 10 mM ammonium formate and
0.1% formic acid and mobile phase B consisting of 90:10
isopropanol:acetonitrile with 10 mM ammonium formate and
0.1% formic acid. The analysis began with 100% A and
increased to 100% B from 1 to 21 min and held at 100% B for 4
min.
MCR-ALS Analysis. For this work, spectral initial estimates
were obtained using IOPA to initiate MCR-ALS. This was
performed using an in-house MATLAB program. MCR-ALS
was also performed using an in-house MATLAB program,
which was based on previously described programs by Allen
and Rutan28 and Bezemer and Rutan.39 The raw three-way LC-
MS data, including the sample mode, are input along with the
initial estimates of the component spectra. The sample
augmentation is performed within the program. The maximum
number of iterations and the convergence criterion are defined
and the constraints are set for the selected components. For the
current work, non-negativity was used in both chromatographic
and spectral modes while selectivity was used in the
chromatographic mode. This selectivity set certain regions in
the chromatographic profiles to zero intensity. Specifically, this
was used at the edges of the retention windows where no peaks
were present to ensure resolution of background signals. A
smoothing constraint based on Eilers’ perfect smoother40 was
investigated for use in the chromatographic mode, but showed
no obvious improvement for these data.
An important consideration when using MCR-ALS is the
degree of rotational ambiguity present in the results. For this
work, the application of constraints along with the selectivity
provided by mass spectrometry minimized rotational ambiguity.
This is supported by the observation that using additional
constraints provided no significant differences in the resolved
analyte profiles.
■
RESULTS AND DISCUSSION
Known Amphetamine Data. To demonstrate the ability
of the proposed algorithm to analyze LC-HRMS data with
MCR-ALS, an analysis of amphetamines (Figure 3) was carried
11096
Article
out. As shown in Figure 3, the peaks between 150 and 190 s
were not selected for analysis. This was due to these analytes,
ephedrine and pseudophedrine, being diastereomers with
identical mass spectra. MCR-ALS requires different spectra to
be able to resolve analyte signals. Otherwise, all of the windows
shown in Figure 3 were analyzed using the methods described
here; however, this discussion primarily focuses on the analysis
of the second window. First, the data were binned to unit mass,
and MCR-ALS was performed. Two components were
observed to contain realistic chromatographic peak shapes.
This agreed with the two known compounds in this retention
time window, PEA and PPA. The resolved mass spectral
profiles are shown in Figure 6A. Three masses were found
Figure 6. Resolved mass spectral profiles for components in window 2
at bin levels of 1 amu (A) and 0.001 amu (B). Below each spectral
plot, the masses corresponding to components 1 (PEA; blue) and 2
(PPA; red) are listed.
above the 5% intensity threshold in this window for both of
these components. These masses are listed in Table 1. These
masses were expanded as described in the Strategy section and
at the 0.1 amu bin level, 3 masses were again determined to be
significant using a 5% intensity threshold.
Table 1. All Extracted Masses for Compounds Resolved
within Window 2 Collected above the 5% Intensity
Threshold at Each Bin Level
bin level PPA PEA
unit 134, 135, 152 105, 106, 122
0.1 134.1, 135.1, 152.1 105.1, 106.1, 122.1
0.01 134.09, 134.10, 134.11, 135.10, 105.07, 105.08, 106.07,
152.10, 152.11 122.09, 122.10,
0.001a 134.096, 135.099, 152.106 105.070, 106.074, 122.096
a
Masses at this level represent the maxima of the spectral peaks.
The masses at this level are more precise than that of the unit
mass bin level. This process was repeated until the 0.001 amu
bin level was reached. It is important to note that at the unit
mass and 0.1 amu bin levels each of the masses were
represented by a single data point (i.e., a spike) in the
spectrum, whereas at the 0.01 and 0.001 amu bin levels, the
peaks are represented by several data points creating a spectral
peak shape, thus more masses were found as significant. At the
final level of resolution in Table 1, only the masses
corresponding to the maximum intensity of each peak are
listed. Because the m/z axis was irregular with intervals at
approximately 0.0015 amu, at the final level of binning (0.001
amu), some bins contained no m/z values. In several instances,
this caused spectral peaks to contain false zero intensities, as
determined by a discontinuous peak. To account for this, a
cubic spline interpolation was performed subsequent to binning
DOI: 10.1021/acs.analchem.6b03116
Anal. Chem. 2016, 88, 11092−11099
Analytical Chemistry Article

but prior to MCR-ALS analysis to ensure a continuous m/z axis

for all spectra in all samples for visualization purposes. The
resolved chromatographic profiles for the two compounds at
the final bin level are shown in Figure 7, with the statistics for

Figure 7. Resolved chromatographic profiles (overlaid samples) for

PEA and PPA at the 0.001 amu bin level.

the final MCR-ALS analysis and calibrations given in Table S2.

The non-Gaussian and jagged characteristics of these peaks are Figure 8. (A) This panel shows the TICs for the total bacterial lipid
caused by injection solvent mismatch41,42 and natural data set. The inset shows the window chosen for analysis. (B) The
fluctuations in electrospray ionization sampling (especially for contour plot shows the LC-HRMS data for the window selected. It can
the highly aqueous mobile phase used for the elution of the be seen that background ions are present in addition to the analytes of
amphetamines), respectively. These characteristics are also seen interest.
in the raw data indicating they are not artifacts of MCR-ALS
processing.
The complexity of the data is evident in both the TIC and
To determine the quantitative performance of this method-
the contour plots of the data, with crowded and overlapping
ology, peak areas were calculated by integrating over the entire
features as well as background attributes, as shown in Figure 8.
resolved chromatographic profiles, shown in Figure 7, and
Analysis with our strategy allowed for the resolution of four
constructing calibration curves. Concentrations were predicted
analyte signals with their respective spectral and chromato-
for each compound in the two test mixtures. The percent errors
graphic profiles shown in Figure 9. As described in the above
in the predictions for the test unknowns are listed in Table 2.
amphetamine results, spline interpolation was employed prior
These errors are all less than 20%, which is comparable to the
to the MCR-ALS analysis at the final bin level. The final MCR-
expected precision of LC-HRMS with no internal standard.
ALS analysis resulted in the compound represented in
Unknown Bacterial Lipid Data. In order to demonstrate
component 1 being split across two components (not shown
the utility of this strategy for untargeted analyses of complex
here). This was clear from the presence of identical mass
samples, this strategy was applied to a bacterial lipid data set
spectra (r2 > 0.99). When added together, the reconstructed
shown in Figure 8. For the purposes of brevity, the results of a
component showed a chromatographic peak shape as seen in
single time window (Figure 8 inset) are reported here.
Figure 9, row 1. This is likely to have been caused by a slight
shift in the mass spectra across the duration of the
Table 2. Percent Errors in the Known Concentrations of chromatographic peak, causing MCR-ALS to resolve the single
Test Samples from the Final Bin Level of 0.001 m/z analyte into two components.
Obtained from This Method The resolved spectral profiles from the final round of analysis
window compound test sample 1 test sample 2 allowed recovery of precise analyte masses which can aid in
1 PE 0.75% 15% compound identification. As seen in Figure 9, the resolved
2 PPA 12% −3.0% spectral profiles in components 3 and 4 showed many
PEA 9.6% -a significant masses. This is due to the relatively low intensities
4 MDA 16% 10% of these analyte signals as seen in their corresponding
Phent 1.1% 11% chromatographic profiles (maximum intensities of 300−400
MDE 15% 4.5% ion counts, as opposed to intensities of 8000 and 2500 counts
Mamp 17% 11% for components 1 and 2, respectively). This reduces the signal-
MDMA 2.6% 9.3% to-(residual) background of the masses corresponding to the
Amp 1.8% -a compounds. Because the spectral profiles are normalized, the
Moxy 6.3% -a residual background ions are intensified in the spectral profiles.
PMMA 4.8% -a Table 3 lists masses believed to be significant in the spectral
5 MTA −0.93% −3.4% profiles of all four components. Also listed in Table 3 are
mCPP 12% -a potential molecular formulas assigned using the LIPID MAPS
6 Bromo 2.6% 4.5% structure search.43 It is interesting to note that in each case, an
7 BP 2.0% 11% isotope peak was recovered, and in the case of compound 4,
Traz −1.6% -a two isotope peaks were recovered. The resolved chromato-
graphic profiles allowed for the determination of retention
a
Blank cells are due to the absence of these compounds from test times as well as relative quantitation between the bacterial
sample 2. strains. While not performed in this work, these resolved
11097 DOI: 10.1021/acs.analchem.6b03116
Anal. Chem. 2016, 88, 11092−11099
Analytical Chemistry Article

■ CONCLUSIONS
A novel method for resolving analyte signals in LC-HRMS data,
which conserves the information from LC-HRMS data, was
developed. In the targeted amphetamine analysis, all known
amphetamine components in each specified window were
recovered using MCR-ALS at every resolution level from unit
mass to 0.001 amu, allowing for the facile quantitation of each
compound. The utility of this procedure is clearly demonstrated
by the finding that the final MCR-ALS step was completed with
as little as 0.55% of the original high-resolution data, with all
relevant high-resolution information being preserved for total
analysis of amphetamine data. The application of this procedure
to unknown, discovery type analyses was also demonstrated
through the analysis of a bacterial lipid data set. The samples
analyzed here were chosen to demonstrate the feasibility of the
proposed strategy; however, this general strategy can easily be
applied to many types of analyses utilizing LC-HRMS.
While a comprehensive comparison between the strategy
outlined in the paper and the ROI approach described
previously was not carried out,23 a brief comparison was
performed in which prediction errors were similar to those
found in Table 2. In order to make a meaningful comparison, a
more comprehensive study should be undertaken with varying
sample conditions. It is our belief that the present strategy will
be more useful in cases of signals with low signal-to-noise. This
is due to the ROI approach requiring a threshold prior to
MCR-ALS analysis, whereas the present strategy only requires a
threshold that is relative to each resolved compound’s most
intense spectral peak. The present strategy also includes fewer
tunable parameters, which may allow for more robust operation
despite requiring significant user interaction. In its current
form, this strategy took up to several minutes per window for
Figure 9. Resolved mass spectral profiles (left column) and complete analysis, including loading data, with user-interaction
chromatographic profiles (right column) of the found components, throughout. While it was not a specific goal of this work, several
labeled 1−4, of the analyzed window at the final bin level. Background
steps of this work would lend themselves well to automation
components were found in the analysis but are not shown here for
clarity. Component 1 required combination of two components as requiring minimal input by the user. Further optimization of
described in the text. the code may also provide significant reduction in analysis time.
These further refinements will allow this strategy to be easily
Table 3. Found Masses in the Bacterial Dataset and Their implemented by analysts with limited chemometrics training.
Associated Molecular Formula
average
retention time associated corresponding molecular formula
■
*
ASSOCIATED CONTENT
S Supporting Information
peak (s) masses (m/z) (within ±0.005 m/z tolerance)a The Supporting Information is available free of charge on the
1 588 211.168 C13H23O2 ACS Publications website at DOI: 10.1021/acs.anal-
212.172 chem.6b03116.
2 589 199.169 C12H23O2
Amphetamine sample table; calibration and fit statistics
200.173
3 592 293.210 C18H29O3 for MCR-ALS (PDF)

■
294.213
447.130 C22H23O10 AUTHOR INFORMATION
448.135
Corresponding Author
4 595 313.214 C21H29O2
314.219
*E-mail: srutan@vcu.edu.
315.193 Notes
a
Calculated from LIPID MAPS structure search 43 The authors declare no competing financial interest.

components can also be submitted to pattern recognition
algorithms such as principal components analysis (PCA) to aid
in distinguishing differences between the bacterial strains.
11098
■ ACKNOWLEDGMENTS
The authors would like to thank the Lipidomics/Metabolomics
Core Facility at Virginia Commonwealth University for the LC-
MS analysis of amphetamines used in this work. The authors
acknowledge financial support from NSF CHE-1507332.
D.W.C. is supported by an Altria Graduate Student Fellowship.
DOI: 10.1021/acs.analchem.6b03116
Anal. Chem. 2016, 88, 11092−11099
Analytical Chemistry Article

■ REFERENCES
(1) Wei, X.; Shi, X.; Kim, S.; Zhang, L.; Patrick, J. S.; Binkley, J.;
(34) Porter, S. E. G.; Stoll, D. R.; Rutan, S. C.; Carr, P. W.; Cohen, J.
D. Anal. Chem. 2006, 78 (15), 5559−5569.
(35) Omar, J.; Olivares, M.; Amigo, J. M.; Etxebarria, N. Talanta
McClain, C.; Zhang, X. Anal. Chem. 2012, 84 (18), 7963−7971. 2014, 121, 273−280.
(2) Tautenhahn, R.; Böttcher, C.; Neumann, S. BMC Bioinf. 2008, 9, (36) Mas, S.; Tauler, R.; de Juan, A. J. Chromatogr. A 2011, 1218
504. (51), 9260−9268.
(3) Danielsson, R.; Bylund, D.; Markides, K. E. Anal. Chim. Acta (37) Golshan, A.; Abdollahi, H.; Beyramysoltan, S.; Maeder, M.;
2002, 454 (2), 167−184. Neymeyr, K.; Rajkó, R.; Sawall, M.; Tauler, R. Anal. Chim. Acta 2016,
(4) Cook, D. W.; Rutan, S. C. J. Chemom. 2014, 28 (9), 681−687. 911, 1.
(5) Smith, C. A.; Want, E. J.; O’Maille, G.; Abagyan, R.; Siuzdak, G. (38) Otto, M. Chemometrics, 2nd ed.; Wiley-VCH: Weinheim, 2007.
Anal. Chem. 2006, 78 (3), 779−787. (39) Bezemer, E.; Rutan, S. C. S. Chemom. Intell. Lab. Syst. 2006, 81
(6) Tsugawa, H.; Cajka, T.; Kind, T.; Ma, Y.; Higgins, B.; Ikeda, K.; (1), 82−93.
Kanazawa, M.; VanderGheynst, J.; Fiehn, O.; Arita, M. Nat. Methods (40) Eilers, P. H. C. Anal. Chem. 2003, 75 (14), 3631−3636.
2015, 12 (6), 523−526. (41) Hsu, S.-H.; Raglione, T.; Tomellini, S. A.; Floyd, T. R.; Sagliano,
(7) Tauler, R. Chemom. Intell. Lab. Syst. 1995, 30 (1), 133−146. N.; Hartwick, R. A. J. Chromatogr. A 1986, 367, 293−300.
(8) Rutan, S. C.; de Juan, A.; Tauler, R. In Comprehensive (42) Jeong, L. N.; Sajulga, R.; Forte, S. G.; Stoll, D. R.; Rutan, S. C. J.
Chemometrics; Brown, S. D., Tauler, R., Walczak, B., Eds.; Elsevier: Chromatogr. A 2016, 1457, 41−49.
Amsterdam, 2009; Vol. 2, pp 249−259. (43) Sud, M.; Fahy, E.; Cotter, D.; Brown, A.; Dennis, E. A.; Glass, C.
(9) de Juan, A.; Jaumot, J.; Tauler, R. Anal. Methods 2014, 6 (14), K.; Merrill, A. H.; Murphy, R. C.; Raetz, C. R. H.; Russell, D. W.;
4964. Subramaniam, S. Nucleic Acids Res. 2007, 35, D527−D532.
(10) Peré-Trepat, E.; Hildebrandt, A.; Barceló, D.; Lacorte, S.;
Tauler, R. Chemom. Intell. Lab. Syst. 2004, 74 (2), 293−303.
(11) Gargallo, R.; Tauler, R.; Cuesta-Sánchez, F.; Massart, D. L.
TrAC, Trends Anal. Chem. 1996, 15 (7), 279−286.
(12) Bortolato, S. A.; Olivieri, A. C. Anal. Chim. Acta 2014, 842, 11−
19.
(13) Peré-Trepat, E.; Tauler, R. J. Chromatogr. A 2006, 1131 (1−2),
85−96.
(14) Pérez, R. L.; Escandar, G. M. Anal. Chim. Acta 2014, 835, 19−
28.
(15) De Llanos, A. M.; De Zan, M. M.; Culzoni, M. J.; Espinosa-
Mansilla, A.; Cañada-Cañada, F.; De La Peña, A. M.; Goicoechea, H.
C. Anal. Bioanal. Chem. 2011, 399 (6), 2123−2135.
(16) Bortolato, S. A.; Arancibia, J. A.; Escandar, G. M. Anal. Chem.
2009, 81 (19), 8074−8084.
(17) Dantas, C.; Tauler, R.; Ferreira, M. M. C. Anal. Bioanal. Chem.
2013, 405 (4), 1293−1302.
(18) Peré-Trepat, E.; Lacorte, S.; Tauler, R. J. Chromatogr. A 2005,
1096 (1−2), 111−122.
(19) Navarro-Reig, M.; Jaumot, J.; García-Reiriz, A.; Tauler, R. Anal.
Bioanal. Chem. 2015, 407, 8835.
(20) Gorrochategui, E.; Casas, J.; Porte, C.; Lacorte, S.; Tauler, R.
Anal. Chim. Acta 2015, 854, 20−33.
(21) Peré-Trepat, E.; Lacorte, S.; Tauler, R. Anal. Chim. Acta 2007,
595 (1−2), 228−237.
(22) Sánchez Pérez, I.; Culzoni, M. J.; Siano, G. G.; Gil García, M. D.;
Goicoechea, H. C.; Martínez Galera, M. Anal. Chem. 2009, 81 (20),
8335−8346.
(23) Tauler, R.; Gorrochategui, E.; Jaumot, J.; Tauler, R. Protoc. Exch.
2015, http://dx.doi.org/10.1038/protex.2015.102.
(24) Bedia, C.; Tauler, R.; Jaumot, J. J. Chemom. 2016, 30, 575−588.
(25) Tauler, R.; Maeder, M.; de Juan, A. In Comprehensive
Chemometrics; Brown, S. D., Tauler, R., Walczak, B., Eds.; Elsevier:
Amsterdam, 2009; Vol. 2, pp 473−505.
(26) Sánchez, F. C.; Massart, D. L. Anal. Chim. Acta 1994, 298 (3),
331−339.
(27) Cook, D. W.; Rutan, S. C.; Stoll, D. R.; Carr, P. W. Anal. Chim.
Acta 2015, 859, 87−95.
(28) Allen, R.; Rutan, S. Anal. Chim. Acta 2012, 723, 7−17.
(29) Sánchez, F. C.; Toft, J.; van den Bogaert, B.; Massart, D. L. Anal.
Chem. 1996, 68 (1), 79−85.
(30) Hugelier, S.; Devos, O.; Ruckebusch, C. J. Chemom. 2015, 29,
448−456.
(31) Parastar, H.; Radović, J. R.; Bayona, J. M.; Tauler, R. Anal.
Bioanal. Chem. 2013, 405 (19), 6235−6249.
(32) Neves, A. C. de O.; Tauler, R.; de Lima, K. M. G. Anal. Chim.
Acta 2016, 937, 21−28.
(33) van Stokkum, I. H. M.; Mullen, K. M.; Mihaleva, V. V. Chemom.
Intell. Lab. Syst. 2009, 95 (2), 150−163.

11099 DOI: 10.1021/acs.analchem.6b03116

Anal. Chem. 2016, 88, 11092−11099