Toward Prediction: Using Chemometrics for the MS of Synthetic Polymers

Transcription

Toward Prediction: Using Chemometrics for the MS of Synthetic Polymers
Anal. Chem. 2010, 82, 8169–8175
Toward Prediction: Using Chemometrics for the
Optimization of Sample Preparation in MALDI-TOF
MS of Synthetic Polymers
Heike Brandt* and Thomas Ehmann
Wacker Chemie AG, Johannes-Hess-Strasse 24, D-84489 Burghausen, Germany
Matthias Otto
TU Bergakademie Freiberg, Institute of Analytical Chemistry, Freiberg, Germany
In recent years, matrix-assisted laser desorption/
ionization time-of-flight mass spectrometry (MALDITOF MS) has become a powerful tool for the study of
synthetic polymers although its mechanism is still not
understood in detail. Sample preparation plays the key
role in obtaining reliable MALDI mass spectra, in
particular, the proper choice of matrix, cationization
reagent, and solvent. There is still no general sample
preparation protocol for MALDI analysis of synthetic
polymers. For known synthetic polymers, such as
polystyrenes and other frequently investigated polymers, application tables in review articles might be a
guide for selecting a MALDI matrix, cationization
reagent, and solvent. For unknown polymers (polymers
which were not analyzed by MALDI-TOF MS before but
whose structures are in part known from the manufacturing process and from NMR analysis as well), the
selection of matrix and solvent is based upon the
polarity-similarity principle. Chemometric methods provide a useful tool for the investigation of sample
preparation because huge data sets can be evaluated
in short time, that is, for extracting relevant information
and for classification of samples, as well. Furthermore,
chemometrics provide a suitable way for the selection
of a proper matrix, cationization reagent, and solvent.
In this paper, a prediction model is presented using
the partial least-squares (PLS) regression. By applying
the model, the suitability of appropriate (nontested)
combinations (matrix, cationization reagent, solvent)
can be predicted for a certain synthetic polymer based
upon the investigation of a few combinations. This
model may help find suitable combinations in a short
time and serve as a starting point for the investigation
of unknown polymers. Results are exemplary presented
for polystyrene PS2850.
It is well-known that the matrix type, cationization reagent,
their concentrations as well as the type of solvent or solvent
* To whom correspondence should be addressed. E-mail: heike.brandt@
wacker.com.
10.1021/ac101526w  2010 American Chemical Society
Published on Web 08/31/2010
mixture, and the spotting technique affect the polymer
distribution.1-16 Many authors described an optimization of
their sample preparation, but to our knowledge, there is no
paper which treats a systematic investigation. Optimization of
sample preparation had been based upon trial and error until
now. Aside from sample preparation, the instrumentation
parameters influence the determined molar mass distribution.17,18
For the investigation of the influence of the instrumentation
parameter designs of experiments (DoE) are the chemometric
methods of choice. Wetzel et al.19 studied significant instrument
parameters for optimization of MALDI-TOF analysis employing
a 25-1 fractional factorial design, while Liland et al.20 had used
a 23 full factorial design and fractional factorial designs.
(1) Weidner, S. M.; Trimpin, S. Anal. Chem. 2008, 80, 349–4361.
(2) Arakawa, R.; Watanabe, S.; Fukuo, T. Rapid Commun. Mass Spectrom. 1999,
13, 1059–1062.
(3) Danis, P. O.; Karr, D. E. Org. Mass Spectrom. 1993, 28, 923–925.
(4) Guttman, C. M.; Wetzel, S. J.; Blair, W. R.; Fanconi, B. M.; Girard, J. E.;
Goldschmidt, R. J.; Wallace, W. E.; Vanderhart, D. L. Anal. Chem. 2001,
73, 1252.
(5) Jackson, A. T.; Yates, H. T.; MacDonald, W. A.; Scrivens, J. H.; Critchley,
G. C.; Brown, J.; Deery, M. J.; Jennings, K. R.; Brookes, C. J. Am. Soc.
Mass Spectrom. 1997, 8, 132–139.
(6) Belu, A. M.; DeSimone, J. M.; Linton, R. W.; Lange, G. W.; Friedman, R. M.
J. Am. Soc. Mass Spectrom. 1996, 7, 11–24.
(7) Deery, M. J.; Jennings, K. R.; Jasieczek, C. B.; Haddleton, D. M.; Jackson,
A. T.; Yates, H. T.; Scrivens, J. H. Rapid Commun. Mass Spectrom. 1997,
11, 57–62.
(8) Hanton, S. D.; Owens, K. G. J. Am. Soc. Mass Spectrom. 2005, 16, 1172–
1180.
(9) Yalcin, T.; Dai, Y.; Li, L. J. Am. Soc. Mass Spectrom. 1998, 9, 1303–1310.
(10) Hoteling, A. J.; Erb, W. J.; Tyson, R. J.; Owens, K. G. Anal. Chem. 2004,
76, 5157–5164.
(11) Meier, M. A. R.; Schubert, U. S. Rapid Commun. Mass Spectrom. 2003,
17, 713–716.
(12) Montaudo, G.; Samperi, F.; Montaudo, M. S. Prog. Polym. Sci. 2006, 31,
277–357.
(13) Rashidzadeh, H.; Wang, Y.; Guo, B. Rapid Commun. Mass Spectrom. 2000,
14, 439–443.
(14) Wang, Y.; Rashidzadeh, H.; Guo, B. J. Am. Soc. Mass Spectrom. 2000, 11,
639–643.
(15) Rashidzadeh, H.; Guo, B. J. Am. Soc. Mass Spectrom. 1998, 9, 724–730.
(16) Trimpin, S.; Keune, S.; Ra¨der, H. J.; Mu
¨ llen, K. J. Am. Soc. Mass Spectrom.
2006, 17, 661–671.
(17) Schriemer, D. C.; Li, L. Anal. Chem. 1997, 69, 4169–4175.
(18) Schriemer, D. C.; Li, L. Anal. Chem. 1997, 69, 4176–4183.
(19) Wetzel, S. J.; Guttman, C. M.; Flynn, K. M.; Filliben, J. J. J. Am. Soc. Mass
Spectrom. 2006, 17, 246–252.
Analytical Chemistry, Vol. 82, No. 19, October 1, 2010
8169
Three questions had to be answered toward prediction:
• Which is the most significant factor of MALDI-TOF
analysis?
• How should the mixing ratio be considered in the modeling process?
• Is it possible to classify the combinations with increasing
suitability for a special polymer?
In a previous paper, we had studied the effect of mixing ratio
employing a 23 full factorial experimental design varying the
volumes of the polymer, matrix, and cationization reagent
solutions.21 An analysis of variance (ANOVA) was used to
investigate the main effects of mixing ratio, matrix type,
cationization reagent, and solvent as well as the two-factor and
three-factor interaction effects of matrix type, cationization
reagent, and solvent. According to Wilks’ lambda of the
multivariate ANOVA statistics, the most significant factor is
the three-factor interaction effect of matrix, cationization
reagent, and solvent; that is, for modeling the selection of
matrix type, cationization reagent, and solvent, the entire
combination has to be considered. Since the mixing ratio affects
the molar mass distribution, as well, the 23 full factorial design
was applied on each studied combination.21 In order to answer
the third question, the MALDI results of the investigated
combinations for a defined polymer were classified by applying
various chemometrics, and the suitability of each combination
was evaluated. For this purpose, ANOVA as well as principal
component analysis (PCA) provided a suitable tool. Applying
these chemometrics, the MALDI results were evaluated and
an overall averaged grade representing the suitability of the
studied combination for the MALDI analysis of the examined
polymer was determined. The three questions toward prediction are answered: the entire combination has to be considered
in the model as well as the mixing ratio, which is varied for
each studied combination using a 23 full factorial design; the
MALDI results could be classified and evaluated.
For modeling, the methods of design and optimization in
organic synthesis by Carlson et al.22,23 were applied, predicting
the suitability of nontested combinations for one polymer using a
partial least-squares (PLS) regression based upon eight training
combinations. For this model, the three componentssmatrix,
cationization reagent, and solventshad to be characterized by
molecular descriptors. The results of our study are exemplary
presented for PS2850, although more synthetic polymers had been
investigated.
EXPERIMENTAL SECTION
Samples and Reagents. Polystyrene PS2850 (PDI ) 1.04)
was obtained from Macherey-Nagel (Du¨ren, Germany). 2-[(2E)3-(4-tert-Butylphenyl)-2-methylpropenylidene]malononitrile (DCTB),
2′,6′-dihydroxyacetophenone (DHAP), 2,5-dihydroxybenzoic acid
(DHB), 2,5-dihydroxy-1,4-benzoquinone (DHBQ), trans-trans-1,4diphenyl-1,3-butadiene (DPBD), dithranol (DT), nicotinic acid
(NA), and 2′,4′,6′-trihydroxyacetophenone (THAP) were used as
(20) Liland, K. H.; Mevik, B.-H.; Rukke, E.-O.; Almøy, T.; Skaugen, M.; Isaksson,
T. Chemom. Intell. Lab. Syst. 2009, 96, 210–218.
(21) Brandt, H.; Ehmann, T.; Otto, M. J. Am. Chem. Soc. Mass Spectrom. 2010,
doi: 10.1016/j.jasms.2010.07.002.
(22) Carlson, R.; Carlson, J.; Grennberg, A. J. Chemom. 2001, 15, 455–474.
(23) Carlson, R. Org. Process Res. Dev. 2005, 9, 680–689.
8170
Analytical Chemistry, Vol. 82, No. 19, October 1, 2010
matrices. Silver trifluoroacetate, lithium trifluoroacetate, copper(II)
trifluoroacetate, lithium chloride, and sodium nitrate were used
as cationization reagents. MALDI matrices and cationization
reagents were purchased from Sigma-Aldrich (Steinheim, Germany) and used as received. The solvents toluene (TOL),
chloroform (TCM), dichlormethane (DCM), methyl ethyl ketone
(MEK), methyl isobutyl ketone (MiBK), and tetrahydrofuran
(THF), stabilized with 2,6-di-tert-butyl-4-methylphenol, were purchased from Merck (Darmstadt, Germany).
MALDI-TOF MS. All experiments were carried out on an
AXIMA Performance MALDI-TOF mass spectrometer (Shimadzu
Biotech, Manchester, U.K.) equipped with a nitrogen laser (λ )
337 nm). The mass spectra were obtained in linear mode with an
accelerating voltage of 20 kV; at least 150 to 200 mass spectra
were averaged for one spectra by scanning a raster. Four
measurements of each sample were carried out for evaluating
standard deviation. Although Wetzel et al.19 showed that detector
voltage can have a high impact on S/N ratios, none of the voltage
settings were manually changed. As shown in numerous papers,
laser energy has a high impact on MALDI-TOF MS analysis. Thus,
in this study, laser energy was optimized for each sample. The
laser energy was set up for each measurement slightly above
threshold of the analyte signal.
Sample Preparation. All combinations of the eight matrices,
five cationization reagents, and six solvents were analyzed (240
combinations). Polymer solutions were prepared at a concentration
of 5 mg/mL, matrix solutions at a concentration of 10 mg/mL,
and cationization reagent solutions at a concentration of 0.1 mol/
L. According to our previous paper dealing with the effect of
mixing ratio, the solutions were mixed in the given volume mixing
ratios of a 23 full factorial design (eight mixing ratios plus the
triplicate center point).21 A 1 µL aliquot of the solution was
hand-spotted on a stainless steel target and allowed to dry by
air. Hardly soluble components in a specified solvent were
made as saturated solutions at ambient temperature and did
not correspond to the above-mentioned concentrations.
Chemometrics. For each of the 240 combinations, the 23 full
factorial design varying the volume mixing ratio was applied
so that more than 2000 mass spectra are available for a
reasonable application of chemometric methods. The employed
chemometrics are not described in this paper; detailed explanations are found in the literature.24-27 For chemometrics, laser
energy, the molecular weight at maximum of distribution (Mp),
signal intensity, signal-to-noise (S/N) ratio, and spectroscopic
resolution (R) were chosen as response variables; the latter
three variables were taken from Mp. Additionally, two response
variables were worked out describing the quality of the
distribution: the number (values between 0 and 4) and intensity
of interfering peaks (caused by clusters or further peak series
besides the principal series; values between 0 and 5) were taken
between Mp and Mp+ 1 repeat units. The higher the values of
the number and the intensity of interfering peaks, the worse
(24) Otto, M. Chemometrics: Statistics and Computer Application in Analytical
Chemistry; Wiley-VCH: Weinheim, Germany, 1999.
(25) Kessler, W. Multivariate Datenanalyse; Wiley-VCH: Weinheim, Germany,
2007.
(26) Brereton, R. G. Applied Chemometrics for Scientists; John Wiley & Sons:
Chichester, UK, 2007.
(27) Eriksson, L.; Johansson, E.; Kettaneh-Wold, N.; Wold, S. Multi- and
Megavariate Data Analysis; Umitrics Academy: Umea, 2001.
Figure 1. PCA scores plot of 70 matrices defined by molecular descriptors (81.97% explained variance by the first and second principal
component). Abbreviations are listed in Table 1.
the mass spectrum. The chosen response variables show
variations in molar mass distributions representing the visual
nature of the spectra and the suitability of the studied combination (matrix, cationization reagent, solvent), as well. The
chemometrics were calculated and evaluated using MATLAB
(R2007b, The MathWorks Inc.) and SOLO (Eigenvector
Research Inc.).
RESULTS AND DISCUSSION
Preface. Since the MALDI mechanism is still not understood
in detail, the modeling of the sample preparation could not include
all possible interactions in the desorption/ionization process. Any
modeling would only be an approximation. The aim of our
modeling is the prediction of the suitability of nontested combinations (matrix, cationization reagent, solvent) based upon a few
tested and evaluated combinations for one single polymer.
Applying the methods of design and optimization in organic
synthesis by Carlson et al.,22,23 a polymer-specific model was
established.
Workflow of Modeling. According to the strategy of Carlson
et al.,22 the following steps toward prediction were carried out:
1.
Applying the dried droplet sample preparation method,
a proper matrix, cationization reagent, and solvent, as
well as the mixing ratio have to be selected. Since the
2.
3.
mixing ratio affects the molar mass distribution for every
studied combination, the whole 23 full factorial design
varying the volume mixing ratio was carried out so that
the experimental space is only made up of the matrix
type, cationization reagent, and solvent.21
The matrix, cationization reagent, and solvent had to be
defined by pertinent descriptors. For the matrices and
cationization reagents, molecular descriptors were applied while, for the solvents, solubility parameters were
more useful. Molecular descriptors were developed for
organic molecules, and we had used them to define the
cationization reagents, too. Since there were not enough
typical characteristics for ionic substances found in the
literature (first, we did not found any data for describing
the trifluoroacetates), the molecular descriptors were also
applied on the cationization reagents. Initially, the solvents were also defined by molecular descriptors, but
during the modeling improvement process, Hansen
solubility parameters had been proved for description
(the polar component δp and the hydrogen bonding
component δh).
Principal component analysis (PCA) was carried out
using the molecular descriptors. The scores of the first
and second principal component were determined for 70
Analytical Chemistry, Vol. 82, No. 19, October 1, 2010
8171
matrices (Figure 1, Table 1) as well as for 29 cationization
reagents (Figure 2). From the score plots, subsets of test
items were selected so that the desired spread of the
experimental space is covered by the chosen test items.
The selected solvents should completely dissolve the
polymer. Figure 3 presents the 2D Hansen solubility map
which provides a feasible tool for choosing appropriate
solvents for a defined polymer.28
4.
Steps 4-7 can be done by a totally mathematical
workflow to select experiments for modeling.22 Since
experiments (combinations) should be chosen which
cover the experimental space (i.e., both well suitable and
suboptimal combinations) for the selection of experiments, a fractional factorial design was created.23 For
the screening design, one test item was chosen of each
quadrant of the PCA score plots as well as four (appropriate) solvents. The created screening design including eight combinations is presented in Table 2.
5.
The eight combinations of the screening design were
analyzed by MALDI-TOF MS. The obtained response
variables were evaluated, and an overall averaged grade
was determined. A grade of 1 describes very well
performing combinations for the studied polymer, while
the grade of 4 represents a failed experiment. Notice that
the whole evaluation is subjectively influenced by the
operator and his aims of research. These eight combinations represent the training data set of the model.
6.
Additionally, further eight combinations had to be
analyzed and evaluated to be used for validation.
7.
With the calibration data set, the model is built using
the partial least-squares regression (PLS1, preprocessing
autoscale, NIPALS (nonlinear iterative partial leastsquares) algorithm). After validation, the model can be
used to predict the suitability of nontested combinations
for the examined polymer.
The matrix X of both data sets consists of 8 rows and 6
columns. The eight training combinations, respectively, the eight
validation combinations make up the rows. The PCA scores of
the first and second principal components of the matrices as well
as the cationization reagents are listed in columns 1-4, whereas
both solubility parameters of the solvents are given in columns 5
and 6. The averaged grades of the suitability of the tested
combinations define vector y. For the prediction of the suitability
of a new combination, the scores of the chosen matrix and
cationization reagent as well as the solubility parameters of the
solvent are needed. For this purpose, the high number of matrices
and cationization reagents was considered in the PCA.
Prediction Results. For PS2850, THF, MEK, DCM, and TOL
were used as solvents for modeling; the results are shown in Table
3. The predicted suitability of the modeled combinations are, aside
from the second combination, consistent with the measured ones.
An absolute mean deviation of ±0.5 was determined for 16
combinations (eight training combinations and eight validation
combinations). Table 4 lists a few predictions applying the PLS1
model for the polar matrix THAP and the nonpolar matrix DCTB
(both matrices were not included in the model). For both matrices,
(28) Brandt, H.; Ehmann, T.; Otto, M. Rapid Commun. Mass Spectrom. 2010,
24, 2439–2444.
8172
Analytical Chemistry, Vol. 82, No. 19, October 1, 2010
Table 1. The 70 Matrices and Their Abbreviations
matrix
abbr.
1,4-bis(5-phenyl-2-oxazolyl)benzene
1,4-dihydroxy-2-naphthoic acid
2-(4-hydroxyphenylazo)benzoic acid
2′,4′,6′-trihydroxyacetophenone
2,5-dihydroxybenzoic acid
2,5-dihydroxy-p-benzoquinone
2′,6′-dihydroxyacetophenone
2-amino-3-methyl-5-nitropyridine
2-amino-4-methyl-3-nitropyridine
2-amino-4-methyl-5-nitropyridine
2-bromo-4,6-dinitroaniline
2-hydroxy-4-methoxybenzoic acid
2-hydroxy-4-methoxybenzophenone
2-mercaptobenzothiazole
2-nitrophenyl dodecyl ether
2-nitrophenyl octyl ether
2-nitrophenyl pentyl ether
2-octyldodecanol
3-aminoquinoline
3-hydroxypicolinic acid
4-benzyloxy-R-cyanocinnamic acid
4-nitroaniline
5-acetylamino-2-mercapto-1,3,4-thiadiazole
5-amino-2-mercapto-1,3,4-thiadiazole
5-chloro-2-mercaptobenzothiazole
5-chlorosalicylic acid
5-ethyl-2-mercaptothiazole
5-methoxysalicylic acid
6-aza-2-thiothymine
6-mercaptopurine
9-anthracenecarboxylic acid
9-bromoanthracene
9-chloroanthracene
9-nitroanthracene
9-phenanthrol
acenaphthene
R-cyano-4-hydroxycinnamic acid
R-cyano-4-phenylcinnamic acid
all-trans retinoic acid
aminopyrazine
anthracene
anthranilic acid
benzo[a]pyrene
caffeic acid
chrysene
dipicolinic acid
dithranol
esculetin
ethylen glycol monosalicylate
ferulic acid
isovanillin
naphthalene
nicotinic acid
norharmane
o-phenanthrolin
o-vanillin
p-cumaric acid
phenanthrene
picolinic acid
pyrazinecarboxylic acid
pyrene
salicylamide
sinapinic acid
terthiophene
tetraethylene glycol dimethyl ether
thiourea
trans,trans-1,4-diphenyl-1,3-butadiene
trans-2-[3-(4-tert-butylphenyl)-2-methyl2-propenylidene]malononitrile
trans-3-indoleacrylic acid
vanillic acid
POPOP
DHNA
HABA
THAP
DHB
DHBQ
DHAP
2,3,5-AMNP
2,4,3-AMNP
2,4,5-AMNP
BDNA
HMB
HMBP
MBT
NPDE
NPOE
NPPE
2-OD
3-AC
3-HPA
CBCA
4-NA
AAMT
AMT
CMBT
5-CSA
EMT
5-MSA
ATT
MP
9-ACA
9-BA
9-CA
9-NA
9-PHOH
ACNP
CHCA
CPCA
RA
APY
ANT
2-AA
BAP
CA
CHR
DPA
DT
DHC
EGS
FA
IVA
NPH
NA
NOR
PHAN
VAN
pCA
PHEN
PA
PZA
PYR
SCA
SA
TTP
TEGD
TU
DPBD
DCTB
IAA
VANA
Figure 2. PCA score plot of 29 cationization reagents defined by molecular descriptors (76.63% explained variance by three principal
components).
AgTFA and Cu(II)TFA should be proper cationization reagents
due to low grades. For DCTB combinations, a slightly better
feasibility is predicted than for THAP. These predictions can
be confirmed by the MALDI results. A very important fact is
also shown in Table 4: the influence of the solubility. Cu(II)TFA
was selected for prediction to show the effect of solvent
selection on the model. For this cationization reagent, THF and
MEK are shown to be good solvents. Hence, the model will
predict Cu(II)TFA combinations to perform well disregarding
the chosen solvents as for chloroform, a nonsolvent for
Cu(II)TFA. Almost all investigated combinations of Cu(II)TFA
using chloroform failed. Solubility is the limiting factor of the
model which is also the reason for the high absolute deviation
of the second combination in Table 3. Taking this fact into
account, the model predicts well and one can choose three or
five or more combinations which are predicted to be suitable and
investigate them.
According to these facts, the combinations of DCTB and THAP
with AgTFA or Cu(II)TFA using THF or MEK as solvents (Table
4) are predicted to be suitable due to their low averaged grade,
and hence, these combinations should yield reliable MALDI mass
spectra of PS2850.
This workflow serves not only for predicting useful combinations but also as a starting point for analysis of unknown polymers;
that is, the eight matrices of the screening design have to be
selected and investigated; the modeling step is omitted. It is
possible to construct a screening design with eight training
combinations for a unknown polymer which has not yet been
characterized by MALDI-TOF MS. On the basis of the production
process, the reactants, the solubility, and, hence, the polarity of
the polymer are known so that a chemical structure can be
assumed. Furthermore, NMR spectra provide useful information.
Feasibility of the Model. The combinations of the screening
design used as calibration experiments are not suitable for all
synthetic polymers. The model (Table 2) predicts well for PS and
PDMS of different molecular masses (recognize that the matrices
and cationization reagents were always the same, but the solvents
were chosen corresponding to the polymer). For other synthetic
polymers, such as poly(vinyl alcohol), or polymers of higher molar
masses, the PLS model has to be adjusted so that other matrices
should be used to satisfy the purpose of step 3.
Protocol of Modeling. The characterization of matrices,
cationization reagents, and solvents by molecular descriptors and
solubility parameters, respectively, has to be calculated once for
a high number of reagents (70 matrices and 29 cationization
reagents were defined by molecular descriptors in this study).
So the investigation of a polymer applying the PLS modeling
workflow starts with the construction of the screening design with
Analytical Chemistry, Vol. 82, No. 19, October 1, 2010
8173
Figure 3. Two-dimensional Hansen solubility map including PS and PDMS and several solvents.
Table 2. Screening Design for Modeling (PS2850)a
combination
QM
1
2
3
4
5
6
7
8
-+-+
++
-+-+
++
matrix
QCR
CR
DHBQ
DPBD
DHB
DT
DHBQ
DPBD
DHB
DT
-+
---+
++
++++
(II)
Cu TFA
AgTFA
AgTFA
Cu(II)TFA
LiTFA
LiCl
LiCl
LiTFA
solvent
THF
TOL
DCM
MEK
MEK
DCM
TOL
THF
a
Q is quadrant of PCA score plots of matrices (M) and cationization
reagents (CR).
Table 3. Evaluated and Predicted Suitability and Their
Deviation of the Training Combinations for PS2850
no.
combination
evaluated
predicted
deviation
1
2
3
4
5
6
7
8
DHAP-Cu(II)TFA-THF
DPBD-AgTFA-TOL
DHB-AgTFA-DCM
DT-Cu(II)TFA-MEK
DHAP-LiTFA-MEK
DPBD-LiCl-DCM
DHB-LiCl-TOL
DT-LiTFA-THF
1
1
2
2
3
4
4
2
0.9
2.5
2.0
1.0
2.6
3.4
3.8
2.3
0.1
–1.5
0.0
1.0
0.4
0.6
0.2
–0.3
the eight combinations for calibration, which were chosen according to the chemical nature of the polymer. Further eight
combinations for evaluation have to be selected. Subsequently,
all of the chosen combinations have to be analyzed for the
examined polymer, and the MALDI mass spectra have to be
evaluated. For the evaluation, the operator has to select several
8174
Analytical Chemistry, Vol. 82, No. 19, October 1, 2010
Table 4. Predicted Suitability of Nontested
Combinations Applying the PLS1 Model for PS2850
combination
predicted values
DCTB-AgTFA-TCM
DCTB-AgTFA-MEK
DCTB-AgTFA-THF
DCTB-Cu(II)TFA-TCM
DCTB-Cu(II)TFA-MEK
DCTB-Cu(II)TFA-THF
THAP-AgTFA-TCM
THAP-AgTFA-MEK
THAP-AgTFA-THF
THAP-Cu(II)TFA-TCM
THAP-Cu(II)TFA-MEK
THAP-Cu(II)TFA-THF
1.9
1.9
1.7
0.9
0.9
0.7
2.1
2.1
1.9
1.1
1.1
0.9
response variables according to his aim of research. When all
combinations are evaluated by grades, the PLS model can be
designed and the suitability of a high number of nontested
combinations can be predicted. The operator will chose the best
predicted combinations (lowest grades) and investigate them. If
a good PLS model was built, these combinations will yield better
results for the selected response variables.
Improvement of Matrix Selection without Modeling. Apart
from its importance for constructing the prediction model,
Figure 1 provides some useful information for sample preparation. On the left side, polar matrices are located, whereas on
the right, nonpolar matrices were found. The bottom right
quadrant contains nonpolar matrices such as polycyclic aromatic hydrocarbons and some sulfur-containing substances.
Nonpolar matrices with polar functionalities are located in the
top right quadrant. NPOE (2-nitrophenyl octyl ether), POPOP
(1,4-bis(5-phenyl-2-oxazolyl)benzene), DCTB, and RA (all-trans
retinoic acid) are matrices that require low laser energies and
are suitable for higher molecular mass polymers. These
matrices are located close to each other. A cluster analysis of
the MANOVA results had shown that DHB and THAP can be
exchanged in most cases, confirming the experimental observations. In Figure 1, both matrices are also located at close
quarters. However, several specific features of the matrices
cannot be captured with the molecular descriptors. So the PCA
score plots of Figure 1 do not provide detailed information of
these features. DHB and THAP could be exchanged in the most
samples in this study, but for some samples, it was not possible.
In these cases, the acidic character of DHB or the nonacidic
character of THAP was crucial for the MALDI analysis.
We had successfully used Figure 1 for a huge number of
experiments because the plot contains more information than the
simple polarity-similarity principle of choice (i.e., the polarity of
the chosen matrix should be similar to the one of the examined
synthetic polymer).
CONCLUSIONS
Applying the PLS1 regression, a polymer-specified model was
built predicting the suitability of nontested combinations. An
absolute deviation of ±0.5 was achieved, predicting the eight
training and the eight validation combinations; that is, a predicted
suitability of 1 ± 0.5 means that this combination should perform
well for the examined polymer. However, a confirmation experiment of a combination predicted to be suitable could fail since
several effects on the obtained molar mass distribution, such as
solubility effects or the occurrence of a precipitate using defined
matrices with silver or copper salts, were not well modeled
employing a small number of training combinations. Nevertheless,
the eight training combinations of the screening design offer the
possibility to be used as a starting point for the analysis of
unknown synthetic polymers.
Received for review June 9, 2010. Accepted August 17,
2010.
AC101526W
Analytical Chemistry, Vol. 82, No. 19, October 1, 2010
8175