Implementation of an Advanced Workflow to Enhancethe Confidence in Compound Identification for Non-

Targeted GC-HR-MS Applications

40th ISCC Symposium, May 29 - June 04, 2016

Eric Dossin, Antonio Castellon, Pierrick Diana, Pavel Pospisil, Mark Bentley, Philippe Guy

Philip Morris International R&D

Reduced-Risk Products (“RRPs”) is the term the company uses to refer to products with the potentialto reduce individual risk and population harm in comparison to smoking combustible cigarettes.

PMI’s RRPs are in various stages of development and commercialization, and we are conductingextensive and rigorous scientific studies to determine whether we can support claims for suchproducts of reduced exposure to harmful and potentially harmful constituents in smoke, andultimately claims of reduced disease risk, when compared to smoking combustible cigarettes.

Before making any such claims, we will rigorously evaluate the full set of data from the relevantscientific studies to determine whether they substantiate reduced exposure or risk. Any such claimsmay also be subject to government review and approval, as is the case in the US today.

2

Outline

3

• Smoke as complex matrix

• GC-HR-MS instrumentation

• Our workflow for complex matrix characterization

• Modeling of Linear Retention Index

• Creation of library with predicted LRI

• Case study (2 examples)

• Conclusion and next steps

PMI Science

4

• PMI is working on various Reduced Risk Products deliveringnicotine containing aerosols

• In this context it is important to fully characterize thechemical composition of these aerosols, in particular aerosolsproduced by heating tobacco compared to smoke fromcigarettes

• For analytical method development purpose, we use areference cigarette (3R4F)

5

TPM - Total Particulate Matter Whole smoke

• Smoke sample generated on a linear smoking machine• The cambridge filter is combined with the impingers

Generation of complex matrix

• Reference cigarette: 3R4F*• Smoking regimen: Health Canada 2 sticks Puff volume: 55mL Puff duration: 2 sec Puff interval: 30 sec Puff count (butt length) 

* University of Kentucky (Kentucky Tobacco R&D Center). http://www2.ca.uky.edu/refcig/

Filter is extracted

2 cold impingers in series

GVP - Gas Vapor Phase

Whole smoke

6

Goal is to screen the broadest range of smoke constituents.Non-Targeted Screening

Analytical Technique: GC-High Resolution (GC-HR-MS)

GC-HR-MS_1(7200A Agilent Q-TOF-MS)

Volatile and semi-volatilesLRI from 500 to 1,900

GC-HR-MS_2(7200B Agilent Q-TOF-MS)

Apolar and polarLRI from 1,000 to 3,000

• Stop when identified: DB‐624 Accurate Mass Library      

(In house PMI Library ≈700 spectra) NIST 14 or Wiley

• Export .cef and review• Purchase of ref std if available

Automated Workflow for Identification in Complex Matrices

7

MassHunterData Acquisition

(EI mode)

DeconvolutionIdentification

MassHunterUnknown Analysis

8x10

0

0.4

0.8

1.2

1.6

+EI TIC Scan 160323_3R4F_004.D

1

Counts vs. Acquisition Time (min)6 8 10 12 14 16 18 20 22 24 26

Acquisition Time (min)6 8 10 12 14 16 18 20 22 24 26

3R4F_Dil 5 (160323_3R4F_004.D)

Cou

nts

8x10

0

0.4

0.8

1.2

1.6

18.6

245

13. 9

501

8.04

61

12. 9

472

23. 7

301

10. 1

395

21. 9

162

6.58

80

19.6

303

17. 0

571

11. 8

957

15. 9

458

24. 8

099

26. 1

756

14. 7

062

20. 6

514

• Stop when identified: DB‐624 Accurate Mass Library Compound library NIST 14 or Wiley

• Export .cef and review• Purchase of ref std if available

8

The present work focuses on building a relevant library containing LRI prediction.

• To predict LRI we used two softwares:1. RapidMiner-Dragon2. ACD/Labs ChromGenius

• To build a relevant LRI prediction system 552 molecules were used:1. Experimental LRI2. Quantitative Structure-Property Relationship (QSPR) and structure similarities

The experimental linear retention indices were randomly split as training (n=401) and test (n=151) sets.Validation set (n=23) confirmed the great performance of both prediction models.

Purpose

Assessment of the Prediction Models (Test Set and Validation Set)

9

LRI p

redi

cted

800

1,200

1,600

1,6001,200800

LRI experimental

RapidMiner values

Manuscript submitted peer‐reviewed journal

1,6001,200800

LRI experimental

800

1,200

1,600

ACD/ChromGenius values

LRI p

redi

cted

Acc

urac

y(A

CD

/Chr

omG

eniu

s vs

. exp

erim

enta

l)1,6001,200800

180%

160%

140%

120%

100%

80%

60%

40%

LRI experimental

Acc

urac

y(R

apid

Min

er v

s. e

xper

imen

tal)

180%

160%

140%

120%

100%

80%

60%

1,6001,200800

40%

LRI experimental

RapidMiner values

R2 : 0.949 R2 : 0.976

ACD/ChromGenius values

n=151 reference standards (test set)n=23 reference standards (validation set)

Based on structure similarity

Based on structural descriptors (QSPR)

10

UCSD is our in-house database that contains 11567 molecules:

7000 chemicals reported as present in tobacco plant and smoke1

3000 molecules associated with flavor properties2,3

1 The Chemical Components of Tobacco and Tobacco Smoke, A. Rodgman, T.A. Perfetti, 2013, 2nd Ed. CRC press.2 Leffingwell, J. C.; Young, H. J.; Bernasek, E. Tobacco flavoring for Smoking Products, R. J. Reynolds Tobacco Company, Winston‐Salem, 1972.3 EFSA flavoring substances database

Unique Compounds & Spectra Database

1013 accurate massspectra and LRI areregistered in UCSD

11

UCSD Web Interface

Creation of the UCSD Compound Library

12

UCSDDatabase (n=11,567)

RapidMiner software

ChromGeniussoftware

Model 2

Model 1

LRI mean value

(within LRI)

Commercial libraryNIST and Wiley (EI)

Extraction of MS spectra

UCSD.xml n = 3,646 molecules

MassHunter Library Editor

• Predicted LRI of flavors and tobacco-related compounds database when thereimplemented in our current identification workflow.• 6,053 molecules were predicted with LRI values between 500 and 1,900• 3,646 molecules have an available nominal EI Mass Spectra (NIST or Wiley)

13

UCSD Compound Library

LRI predictions were associated with the 3,646 nominal EI mass spectraextracted from commercial libraries.

LRI Prediction

• Stop when identified: DB‐624 Accurate Mass Library(In house PMI Library≈700 spectra)

NIST 14 or Wiley• Export .cef and review• Purchase of ref std if available

Different Output With the New Workflow

14

Data Acquisition

(+EI)

DeconvolutionIdentification

MassHunterUnknown Analysis

Software

Implementation of predicted LRI allows alternative proposals.

•

••••

••

•

•••

••

••••••••

••

• Stop when identified: DB‐624 Accurate Mass Library(In house PMI Library≈700 spectra)  UCSD Library (LRI predicted) NIST 14 or Wiley

• Export .cef and review• Purchase of ref std if available

1st Example / Evaluation

15

14.50 15.00 15.50 16.00

16

.36

7

15

.14

5

15

. 94

6

15

. 43

3

14

.70

6

15

.58

8

15

.78

4

14

.94

6

14

.52

4

16

.08

61

6.2

22

16

.47

1

Acquisition Time (min)13.00 14.00 15.00 16.00 17.00 18.00

3R4F_Rep 4 (160323_3R4F_004_duplicate.D)

Co

un

ts

7x10

0

0.5

1

1.5

2

2.5

3

3.5

4

4.5

5

13

.95

0

12

.94

7

16

.91

6

13

.82

6

17

.86

5

12

.79

5

16

.36

7

15

.14

5

15

.94

6

12

.54

2

17

.57

7

15

.43

3

13

.13

0

17

.36

3

14

.70

6

13

.39

9

14

.20

9

18

.15

2

LRI1204.02

LRI1214.03

3-Methyl-1H-indene

Smoke sample

Confirmation

1st Example / Evaluation

16

14.50 15.00 15.50 16.00

16

.36

7

15

.14

5

15

. 94

6

15

. 43

3

14

.70

6

15

.58

8

15

.78

4

14

.94

6

14

.52

4

16

.08

61

6.2

22

16

.47

1

Acquisition Time (min)13.00 14.00 15.00 16.00 17.00 18.00

3R4F_Rep 4 (160323_3R4F_004_duplicate.D)

Co

un

ts

7x10

0

0.5

1

1.5

2

2.5

3

3.5

4

4.5

5

13

.95

0

12

.94

7

16

.91

6

13

.82

6

17

.86

5

12

.79

5

16

.36

7

15

.14

5

15

.94

6

12

.54

2

17

.57

7

15

.43

3

13

.13

0

17

.36

3

14

.70

6

13

.39

9

14

.20

9

18

.15

2

LRI Prediction

Alternative hits proposal

LRI1204.02

LRI1214.03

1st Example / Confirmation

17

Same RT Similar EI Mass Spectra

2-Methyl-1H-indeneCommercial

Smoke sample

14.50 15.00 15.50 16.00

16

.36

7

15

.14

5

15

.94

6

15

. 43

3

14

.70

6

15

.58

8

15

.78

4

14

.94

6

14

.52

4

16

.08

61

6.2

22

16

.47

1

Acquisition Time (min)13.00 14.00 15.00 16.00 17.00 18.00

3R4F_Rep 4 (160323_3R4F_004_duplicate.D)

Co

un

ts

7x10

0

0.5

1

1.5

2

2.5

3

3.5

4

4.5

5

13

.95

0

12

.94

7

16

.91

6

13

.82

6

17

.86

5

12

.79

5

16

.36

7

15

.14

5

15

.94

6

12

.54

2

17

.57

7

15

.43

3

13

.13

0

17

.36

3

14

.70

6

13

.39

9

14

.20

9

18

.15

2

3-Methyl-1H-indene is confirmed

2-Methyl-1H-indene is confirmed1214.31 0.282

LRI1204.02

LRI1214.03

81.32177‐47‐1

Evaluation of NIST Proposals

18

The combination of mass spectral similarity and LRI modeling enhances theconfidence in compound identification.

Alternative ranking using the LRI model!!!

19

2nd Example / Evaluation

Acquisition Time (min)11.00 11.50 12.00 12.50

Co

un

ts

6x10

0

1

2

12

.79

5

11

.64

8

11

.89

6

11

.43

4

12

.54

2

12

.2

25

12

.63

3

11

.12

9

12

.01

5

11

.76

35

12

.40

8

12

.15

1

11

.84

27

12

.07

5

12

.70

2

12

.29

1

11

.54

4

11

.01

8

11

.21

2

11

.33

7

12

.35

1

3R4F_Rep 4 (160323_3R4F_004_duplicate.D)

Benzene,1-ethyl-4-methyl

3R4F

As Match Factor > 90%, the confidence in identification is very high.

LRI992.25

Let’s go further to identify this co-eluting compound!!!

20

Again, the combination of mass spectral similarity and LRI modeling improvesthe identification process

Alternative ranking using the LRI model !!!

Already parts of our accurate mass library

Evaluation of NIST Proposals

21

Acquisition Time (min)11.00 11.50 12.00 12.50

Co

un

ts

6x10

0

1

2

12

.79

5

11

.64

8

11

.89

6

11

.43

4

12

.54

2

12

.2

25

12

.63

3

11

.12

9

12

.01

5

11

.76

35

12

.40

8

12

.15

1

11

.84

27

12

.07

5

12

.70

2

12

.29

1

11

.54

4

11

.01

8

11

.21

2

11

.33

7

12

.35

1

3R4F_Rep 4 (160323_3R4F_004_duplicate.D)

Benzene,1-ethyl-4-methyl

3R4F

Benzene,1-ethyl-3-methyl

Final Confirmation

Both ethyl-toluene isomers wereconfirmed in the smoke sampleexplaining the co-eluting peaks.

22

Conclusion & Next Steps

Implement the LRI prediction in the MassHunter Unknown Analysis software

Dossin, E. et al. Prediction Models of Retention Indices for Increased Confidencein Structural Elucidation during Complex Matrix Analysis: Application to GasChromatography Coupled with High Resolution Mass Spectrometry coming inAnal. Chem. in 2016

• Combination of state-of-the-art instrumentation with advanced chemoinformatictools (Predicted LRIs using QSPR methods) enhance the confidence level incompound identification

• This combined approach reduces the list of putative compounds to purchase forfinal confirmation, leading to

• Shortened time for compound identification• Reduced total cost for ordering chemicals• Reduced rate of false positive compound identification

• Automation of the workflow reduces the time for complex matrix characterization

Implementation of an Advanced Workflow to Enhance the Confidence in Compound Identification for Non-Targeted GC-HR-MS Applications

40th ISCC Symposium, May 29 - June 04, 2016

Eric Dossin, Antonio Castellon, Pierrick Diana, Pavel Pospisil, Mark Bentley, Philippe GuyPhilip Morris International R&D

THANK YOU eric.dossin@pmi.com

24

GC-HR-MS_2(7200B Agilent Q-TOF-MS)

Goal is to screen the broadest range of smoke constituents

Apolar & PolarLRI from 1000 to 3000

Column HP-5MS (30m, Ø=0.25mm)5% phenyl‐arylene95% dimethyl‐polysiloxane

Carrier gas:  He

Injection volume: 0.5 µL

Injection mode split 5:1

Sample spiked with several deuterated IS 

Full scan acquisition mode (m/z 40‐620 amu)

BFTSA derivatization (labile hydrogens, alcohols, carboxylic acids, amides…)

GC-HR-MS_1(7200A Agilent Q-TOF-MS)

Volatile and semi‐volatilesLRI from 500 to 1900

Column DB-624 (30m, Ø=0.25mm) 6% cyanopropyl‐phenyl94% dimethyl‐polysiloxane

Carrier gas:  He

Injection volume: 250 µL (headspace)1 µL (liquid)

Injection mode split 5:1

Full scan acquisition mode (m/z 20‐500 amu)

Analytical Technique: GC-High Resolution (GC-HR-MS)

Accuracy Data of Predicted vs Experimental LRI Values

25

Accu

racy

(AC

D/C

hrom

Gen

ius

vs. e

xper

imen

tal)

1,6001,200800

LRI experimental

180%

160%

140%

120%

100%

80%

60%

40%

Accu

racy

(Rap

id M

iner

vs.

exp

erim

enta

l)

180%

160%

140%

120%

100%

80%

60%

1,6001,200800

LRI experimental

40%

RapidMiner ACD/ChromGenius

n=151 reference standards (test set)n=23 reference standards (validation set)

C5H5NOC6H10

C2HBrClF3

C2H3N C12H20O7

C5H10OC6H14O C7H6O2

Manuscript in revision

r2 : 0.949r2 : 0.976

26

Mathematical Formula

Exp =

Exp+ ( Pred‐ Exp)2