Dept seminar

1
Chemometric Methods for GC x GC
LCDR Gregory J. Hall Glenn S. Frysinger
Department of Science U.S. Coast Guard
Academy New London, Connecticut gregory.hall_at_uscga
.edu
2
LCDR Gregory J. Hall
1995 B.S. Marine Science U.S. Coast Guard
Academy 1995 1997 Operations Officer, USCGC
SPAR 1997-1998 M.S. Chemistry, Tufts
University 1998-2000 Rotating Military Faculty,
USCGA 2000 Appointed to the PCTS 2002 2004
Ph.D. sabbatical, Tufts University 2006 Ph.D.
Chemistry, Tufts University Chemometric
Characterization and Classification of Estuarine
Water through Multidimensional Fluorescence
3
Permanent Commissioned Teaching Staff (PCTS)
About 23 officers ranked from LT to CAPT Provide
the interpreters between the military and
civilian faculty and leadership for the
college Teaching, Service, and Scholarship
expected Ph.D. required
4
LCDR Gregory J. Hall
5
What IS Chemometrics?
Chemometrics is the chemical discipline that uses
mathematical, statistical and other methods
employing formal logic to design or select
optimal measurement procedures and experiments,
and to provide maximum relevant chemical
information by analyzing chemical data. (D.L.
Massart Chemometrics, Elsevier, NY,1988)
6
Chemometrics already covered and to come

1. Difference Chromatograms
2. Property Modeling
3. Clustering
4. Chromatograph Prediction
5. Mass Spec searching
6. Template Construction
7. XICs
8. Retention Indices

You are all already chemometricians!
7
Today

• Data Structures How I view GC x GC data
• Variance - PCA
• Classification SIMCA, PCR-DA
• Regression PLS
• Peak Resolution - PARAFAC
• Preprocessing Alignment
• The way forward, humble opinions

8
Data GC x GC - FID
Second Dimension
Intensity Values
I
Second Dimension
J
X
sample
K
First Dimension
First Dimension
Dataset Data Object
Chromatogram Stack Three way 4 Dimensions
Chromatogram Two way 3 Dimensions
9
Data GC x GC -TOF
Second Dimension
m/z
First Dimension
Sample (Date?)
Dataset Four way 5 Dimensions !
X
10
Principal Components Analysis (PCA)
PC 1
i
PC 2
j
variable 3
Q
T2
variable 1
variable 2
11
Principal Components Analysis (PCA)
P
components

E

T
Samples
data
model
residuals
Goal - Variance capture
12
Multi-way Principal Components Analysis (MPCA)
Our data 15 x 410,000
Wise, B. M. Gallagher, N. B. Bro, R. Shaver,
J. M. Windig, W. Koch, R. S. PLS Toolbox 4.0
Eigenvector Research, Inc. Wenatchee, WA, 2006.
13
GC GC/MS TIC of Fire Debris
4.0
3.0
Time (s)
2.0
1.0
0.0
Time (min)
6 clean carpet samples 5 gasoline
samples 6 doped carpet samples
14
PCA Model Specifics

• Only two carpet classes included
• 4 PCs 98 variance
• Two random samples per class left out, all
  gasoline samples left out of training set
• Left out samples projected onto the model later.

15
PC 1 - Loadings
Red positive loadings, correlated Blue
negative loadings, anti-correlated
16
PC 2 - Loading
Chemically interpretable results! Next step -
classification
17
Principal Components RegressionDiscriminant
Analysis (PCR-DA)
Y
w/ accelerant wo/ accelerant
0 1
0 1
0 1
0 1
1 0
1 0
1 0
1 0
1 0
PC 1
i
PC 2
j
Q
variable 3
T2
variable 1
variable 2
18
Regression Vector
Red positive loadings Blue negative loadings
19
Regression Vector Zoom
150
100
30
20
25
20
Principal Components Regression Predictions
Sample Scores on the Regression Vector
1.8
1.6
Gasoline
1.4
1.2
Arson Debris
1
0.8
0.6
0.4
Unaltered Carpet
0.2
0
-0.2
1
6
7
12
17
Discriminant Analysis 1 Member of Arson Class
21
Classification Soft Independent Model of Class
Analogy (SIMCA)
PC 1
i
PC 2
j
variable 3
Q
T2
variable 1
variable 2
22
SIMCA Model Specifics

• PCA modeled for 2 classes Arson , not Arson
• Each model had 2 PCs with 99 variance captured
• One random samples per class left out, all
  gasoline samples left out of training set
• Left out samples projected onto each model
  later.

23
Arson Case SIMCA Results
1
1
In Doped Class
In Carpet Class
0
0
Carpet Samples
Carpet
Doped
Gasoline
Carpet
Doped
Gasoline
Carpet Test
Doped Samples
Doped Test
Gasoline Test
1
2
Not in any Class
Nearest Class
0
1
Carpet
Doped
Gasoline
Carpet
Doped
Gasoline
24
Arson Case SIMCA Fit Statistics
Fit Statistics for Doped Carpet Class
Fit Statistics for Carpet Class
0.25
1
0.2
0.8
0.15
0.6
T2 Residuals
T2 Residuals
0.1
0.4
0.05
0.2
Carpet Samples
0
0
Carpet Test
0
50
100
150
200
250
0
500
1000
Doped Samples
Q Residuals
Q Residuals
Doped Test
Gasoline Test
0.03
0.04
0.025
0.03
0.02
T2 Residuals
T2 Residuals
0.02
0.01
0.015
0
0.01
-0.01
-4
-2
0
2
4
6
8
-10
0
10
20
30
Q Residuals
Q Residuals
25
Parallel Factor Analysis (PARAFAC)
I
I
B
G
R
C
I


K
R
X
J
E
J
A
J
K
K
R
c1
c2
c3
I
I
b1
b2
b3


E


J
X
J
a1
a2
a3
K
K
26
Parallel Factor Analysis (PARAFAC)
Score
Sample
Factor 1
Loading
Loading
I
PARAFAC
First Dimension
Second Dimension
J
X
K
Score
c2
GC x GC - FID Chromatogram Stack
b2
Sample
Factor 2
Loading
Loading
a2
First Dimension
Second Dimension
27
Parallel Factor Analysis (PARAFAC) GC x GC - TOF
Sinha, A. E. Fraga, C. G. Prazen, B. J.
Synovec, R. E. Journal of Chromatography A 2004,
1027, 269-277.
28
Parallel Factor Analysis (PARAFAC)
Score
m/z
Factor 1
Loading
Loading
I
PARAFAC
First Dimension
Second Dimension
J
X
K
Score
c2
GC x GC - TOF Sample
b2
m/z
Factor 2
Loading
Loading
a2
First Dimension
Second Dimension
29
Parallel Factor Analysis (PARAFAC) GC x GC - TOF
Complex Environmental Sample
Sinha, A. E. Fraga, C. G. Prazen, B. J.
Synovec, R. E. Journal of Chromatography A 2004,
1027, 269-277.
30
PARAFAC Results
Sinha, A. E. Fraga, C. G. Prazen, B. J.
Synovec, R. E. Journal of Chromatography A 2004,
1027, 269-277.
31
PARAFAC Results
Sinha, A. E. Fraga, C. G. Prazen, B. J.
Synovec, R. E. Journal of Chromatography A 2004,
1027, 269-277.
32
GCImage screen capture
GC GC/MS Peak Deconvolution PARAFAC?
NIJ0221 100 µg 75 Wx gasoline / nylon carpet
matrix
33
Partial Least Squares (PLS)
P
variables
latent variables


T
samples
E
properties
F

Q
Y

samples
T
data
model
residuals
34
PLS Results Naphthalenes in Jet Fuel
Johnson, K. J. Prazen, B. J. Young, D. C.
Synovec, R. E. Journal of Separation Science
2004, 27, 410-416.
35
Alignment Strategy 1

• Experimental Design

Alignment Strategy 2

• Templates / Peak Tables

Alignment Strategy 3

• Retention Index

36
Alignment Strategy 4
Piecewise Correlation Maximization
Pierce, K. M. Wood, L. F. Wright, B. W.
Synovec, R. E. Analytical Chemistry 2005, 77,
7735-7743.
37
Alignment Strategy 5

• Warping

Kaczmarek, K. Walczak, B. de Jong, S.
Vandeginste, B. G. M. Journal of Chemical
Information and Computer Sciences 2003, 43,
978-986.
38
Alignment Strategy Proposal 1

• Anchor Warping

39
Alignment Strategy Proposal 1

• Anchor Warping

40
Alignment Strategy Proposal 2

• DTW Piecewise Hybrid

2nd Dimension Piecewise
1st Dimension DTW Alkanes?
41
Humble Opinions

1. GC x GC is tremendously interesting data
2. Tremendous amounts of work possible, even with
   data that presently exists. Good alignment will
   open up even more possibilities
3. Include the Chemist in the analysis
4. Include the Chemometrician in the experimental
   design

42
Future?

• More PCA, PCR, PLS, PARAFAC
• Regression certainty calculations
• NPLS, NPLS-DA
• 4. Holistic, automatic alignment strategies
• 2D COW or DTW ?
• PARAFAC 2 ?
• 5. User driven alignment strategies
• Anchor warping
• 6. Inclusion on m/z axis
• Purity, CODA?

43
Acknowledgements
U.S. Coast Guard Academy Alexander Trust You all!