mass

---

340

Anal. Chem. 1 9 ~ 557, , 348-352

the volatile trimethyl borate by codistillation with methanol, and the residue was heated 2 h at 85 "C, in vacuo to completely convert the aldonic acids into aldonolactones. Lower temperatures and/or shorter time periods resulted in partial conversion and the isolation of some free aldonic acids. Complete conversion is necessary because aldonic acids will not form N-propylaldonamide when treated with l-propylamine. Even though the amine is very basic, it did not cause epimerization under the conditions studied. Acetylation with pyridine and acetic anhydride requires 1h and is quantitative. Dehydration reactions occur with the lactone (16) but not with the amide. Essentially the original aldoses have been converted into alditol acetates, and the original aldonic acids and aldonolactones have been converted into corresponding N-propylaldonamides. Calibration curves constructed by plotting detector response vs. milligrams of sample (1.2-6.7 mg) have a correlation coefficient greater than 0.995. A wider range of sample sizes can be accommodated by suitable concentration or dilution of the injected sample. There are a number of advantages to the new procedure. The method is able to handle complex mixtures of aldonic acids and aldoses. Alduronic acids can be determined by suitable modifications. Because the carboxyl group is kept, the stereochemical configuration of the sugar is maintained, unlike the differential method (2)wherein the lactone is reduced to the alditol. The new method adds a degree of versatility not found in most methods. The N-propyl group can be substituted easily by an N-butyl, N-amyl, or even an N-hexyl by replacing 1-propylamine by some other amine. The lactone will react under basically the same conditions to give a homologous derivative. The advantages are readily apparent in that the larger the N-alkyl group, the greater the elution time relative to the alditols or other possible compo-

nents, e.g., phenyl P-D-ghcopyranoside. In fact, N-propylxylonamide coelutes with phenyl 0-D-glucopyranosides;by changing to N-amylamine, one is able to separate them into two peaks while maintaining the resolution seen in Figure 2. Registry No. Ribonic acid, 17812-24-7; mannonic acid, 6906-37-2;gluconic acid, 526-95-4; galactonic acid, 13382-27-9; L-arabinose, 5328-37-0; D-xylose, 58-86-6;D-mannose, 3458-28-4; D-glucose, 50-99-7; D-galactose, 59-23-4; myo-inositol, 87-89-8; ~-mannono-1,4-lactone,22430-23-5; D-galactono-1,4-lactone, 2782-07-2; D-g~ucono-1,5-~actone, 90-80-2. LITERATURE CITED (1) (2) (3) (4)

Humphrey, A. E.; Reiiiy, P.J. Biotechnol. Bioeng. 1965, 7,229-243. Lehrfeld, J. Anal. Biochem. 1981, 115, 410-418. Sloneker, J. H. Methods Carbohydr. Chem. 1972, 6 , 20-24. Easterwood, J. M.; Huff, B. J. L. Sven. fapperstidn. 1969, 72, 768-772.

(5) Dmitriev, B. A.; Backinowsky, L. V.; Chizhov, 0. S.; Zoiotarev, 8. M.; Kochetkov, N. K. Carbohydr. Res. 1971, 19, 432-435. (6) Lehrfeld, J. Anal. Chem. 1984, 5 6 , 1803-1806. (7) Blake, J. D.; Richards, G. N. Carbohydr. Res. 1968, 8 , 275-281. (8) Blake, J. D.; Richards, G. N. Carbohydr. Res. 1970, 14, 375-387. (9) Morrison, I. M.; Perry, M. B. Can. J. Biochem. 1968, 44, 1115-1126. (10) Sawardeker, J. S.; Sloneker, J. H.; Jeanes, A. Anal. Chem. 1965, 37, 1602-1604. (11) AbdeCAkher, M.; Hamilton, J. K.; Smith, F. J. Am. Chem. SOC.1951, 73,4691-4692. (12) Sjostrom, E.; Haglund, p.; Janson, J. Acta Chem. Scand. 1968, 2 0 , 17 18-17 19. (13) Brown, H. C. "Hydroboration"; W. A. Benjamin: New York, 1962; p 243. (14) "Sodium Borohydride"; ThiokoilVentron: Danvers, MA, 1979; p 4. (15) Lee, J. B. Chem. Ind. 1959, 1455-1456. (16) Nelson, C. R.; Gratzl, J. S. Carbohydr. Res. 1978, 6 0 , 267-273.

RECEIVED for review June 25,1984. Accepted October 9,1984. The mention of firm names or trade products does not imply that they are endorsed or recommended by the U.S.Department of Agriculture over other firms or similar products not mentioned.

Analysis of Complex Mixtures by Gas Chromatography/Mass Spectrometry Using a Pattern Recognition Method Mingjien Chien

Givaudan Corporation, Clifton, New Jersey 07014

Analysis of a complex mlxture Involves a gas chromatographlc separation followed by a mass spectral search for Identification of each component. Thls Is an extremely tedlous and time-consumlng process for fragrances and flavors whlch sometimes contaln hundreds of components. Thls presentatlon describes a computer pattern recognttlon method for the rapid ldentlficatlon of components in a complex mixture, the comparlson between the complete GC profiles of different samples, and the recognltlon of a group of compounds as a single entity. The central part of thls method is a comparlson algorlthm which compares the components In two mixtures uslng the K nearest neighbor classification rule. I t is Illustrated by comparing essential oils of dlfferent sources and by detectlng essentlal 011s In perfumes. A varlety of procedures deslgned to expedlte the search process are also dlscussed.

In the past decade, with the development of automated computer processes, gas chromatography/mass spectrometry

has become one of the most powerful tools in the analysis of complex mixtures. Today, computer-processed data acquisition, quantitation, and consequently the search of a data base for each peak have become a routine practice. However, for a sample containing no less than 100 components, the conventional approach of an exhaustive mass spectral search for each GC peak is still very tedious and time-consuming. Besides, mere peak identification of such complex mixtures often provides us with information of little value. On the other hand, there is indeed much valuable information concealed under the complicated GC pattern and can be brought out only through more sophisticated data processing. In the analysis of flavor and fragrance meterials, complex mixtures are routinely encountered. However, most components in the sample are very likely to be known compounds. When the same type of flavor or fragrance is analyzed, virtually the same set of compounds is found each time. Complete mass spectral search for identification is often unnecessary. Furthermore, many components in the mixture may have originated from a single source, such as an essential oil.

0 1984 American Chemical Society 0003-2700/85/0357-0348$01.50/0

ANALYTICAL CHEMISTRY, VOL. 57, NO. 1, JANUARY 1985

If most components can be categorized into different groups, the complexity of such samples will be much reduced. Accordingly, three features of interest are (a) to develop a rapid method of identification of components without complete mass spectral search, (b) to establish a similarity parameter between samples by comparing their GC profile, and (c) to recognize a group of components in the mixture as a single entity. This paper presents a computer pattern recognition method developed in our laboratory for these purposes. The principal part of this method is a comparison algorithm based on the K nearest neighbor classification rule. It is illustrated by comparing essential oils from different sources and detecting essential oils in complex mixtures. This method should be also useful in the study of many other types of mixtures, such as biomedical fluids (I,2),environmental pollutants (3-5), and natural products (6). THEORY

Computer pattern recognition has long been used as a method to aid the analysis of spectroscopic data (69). Among others, the K nearest neighbor classification method has the advantage of higher predictive ability and of being simpler in both concept and computation (10, 12). A multicomponent system is represented as a point in a multidimensional space, with each component being a dimension. The Eucliedean distance between two systems is expressed in eq 1 N

d = [C (C1i - C2i)2]1/2 i=l

(1)

where N is the total number of components and C1, and Czi are the concentrations of component i in systems 1 and 2, respectively. This distance directly reflects their dissimilarity. An unknown system is, therefore, identified according to its nearest neighbors. T o present the results in a more conceptually understandable term, two functions F12and F2, between two systems containing nl and n2 components, respectively, were defined F12

=

F21

=

where W(s are weighing factors, P is the number of common components found in both systems, and D1and D2are the greatest possible distances between them. These two equations are modifications of a relation defined by Kowalski and Bender (12, 13). Let CI1’sand C2i’s be expressed in a relative scale from 1 to 1000. D 1and D2 then become n1

D, =

c Wi(l0g 1000)2

i=l

(4)

n2

Dz = C Wi(l0g 1000)2 i=l

(5)

By definition, F12measures the probability for system 1as part of system 2. If most components in system 1can be found in system 2 at similar levels, F12will be close to 1OOO. If system

349

1contains many extra components not accounted for in system 2, Flz will be considerably lower. The same principle holds true for F21. EXPERIMENTAL SECTION

Apparatus. GC/MS data are acquired on a Finnigan 4000 mass spectrometer equipped with an INCOS data system. A Hewlett-Packard cross-linked OVlOl fused silica column is used for all experiments. Oven temperature is programmed from 50 to 210 “C at a rate of 4 “C/min without initial holding time. All programs are written in FORTRAN IV, using the NOVA 3 computer of the Finnigan data system. These programs are compatible to the MSDS mode in INCOS. Procedure. Four major steps are involved from the acquisition of data to the final characterization of the sample mixtures. Step 1. After GC/MS data are acquired in a usual manner, a sample file containing the retention time, peak area, and m l e values of four selected mass ions (referred to as base peaks) of each GC peak is created. The base peaks are the four most intense ions that are at least 3 amu apart from each other. Step 2. All retention times in the sample file are converted into Kovats retention indexes based on normal hydrocarbons, according to the polynomial equation K = CAi(log T)i

(6)

i

where Tis the retention time and K is the Kovats retention index. In our programs, a polynomial function of the order of 20 i$ used. The coefficients Ai are determined by the least-squares fitting of hydrocarbon data. Step 3. The Kovats retention index and the four selected mass ions of each GC peak are used for identification by searching a library file. This library file, constructed in our laboratory, contains retention times and selected mass ion information of many common substances used in the fragrance and flavor industry. Both sample and library files are arranged in ascending order of Kovats retention indexes. In the search procedure, the Kovats index and base peak information of each entry in the sample file is compared with each reference compound in the library file. When the difference in Kovats indexes is equal or less than 5 units and the four base peaks are matched, a positive identification is considered to be achieved. Search procedure continues until every entry in the sample file is compared. Since both sample and library files are sequentially arranged, the identification of the entire sample file can be accomplished in a single step of scanning through both files simultaneously. It is not necessary to search the entire library for each entry in the sample file. Step 4. After the componentsin sample mixtures are identified, it is possible to determine the likeness between two profiles by calculating functions Fl, and F,, according to eq 2 and 3. One example is to compare a sample profile to a reference profile in a library in order to determine its authenticity. The result is visualized in a bar graphic pattern (Figures 1-3) in which GC peaks are represented by vertical lines with their intensities plotted against Kovats retention indexes. To determine the presence of an essential oil in the sample, a seearch program that scans through an essential oil library is used. The sample profile is compared to each reference profile in the library followed by a calculation of FZl,where 1 and 2 refer to sample and reference, respectively. A high F2, for a particular essential oil indicates its presence in the sample. However, to search the entire library containing a few hundred entries is extremely time-consuming. To expedite this search process, two prefiltering procedures designed to eliminate references which are unlikely to be a good match are used. The details of these prefiltering procedures will be discussed later in this section. Construction of Essential Oil Library. The library consists of individual reference files for each essential oil and a directory table that links all reference files together. Data of an authentic sample of an oil are acquired and processed in the same manner described in steps 1-3. After identification of its components, a reference file of this oil, having the same format of the sample file, is generated and added into the library. An index table containing the information of 10-15 selected “characteristic”

350

ANALYTICAL CHEMISTRY, VOL. 57, NO. 1, JANUARY 1985 29

59

CERINlUM

CARROT SEED

EGypnuI. WERM

F,t :841

I

2

I

GERANIUM

BOURBON & T U . WERER

22

I1

: 908

F,,

:644

I

CARGO1 SEED G I V LOlll 9037-83

F,,

F,,

: 876

,,8

],I,:,

I 600

Figure 1. Comparison of GC profiles of two types of geranium oils. Each line in the bar graphic pattern represents a GC peak. 13

Fa,

: 621

F,,

: 875

9

I

I

31

CARDANON GIV WTB 4031-83

I4

I

I440

2888

components is also created. These characteristic components, referred to as index peaks, are used in the prefilter procedures. All data in the reference file are carefully examined. Errors, if any, are corrected by an editing program. Prefiltering Procedures. The first prefilter is a pTeliminary comparison to eliminate those references in the library whose principle constituents are not found in the sample. For example, patchouli oil will be eliminated if patchouli alcohol, its largest component, is not found in the sample. For each of the remaining references, the second prefilter reads the 10-15 index peaks and finds their corresponding concentration in the sample. F,, is then calculated based on these values. After the entire library has been scanned, a list of essential oils is printed according to the descending order of Fzl. RESULTS AND DISCUSSION Indentification of Components i n Mixtures. The use of only retention indexes and four mass ions, instead of the complete mass spectrum from commercial libraries, in matching unknown and reference compounds greatly expedites the identification process with little effect on its accuracy. The detection limit is also greatly reduced, since smaller mass ions need not be seen. Much time is also saved since it is not

j,(, 1 1 1 ~

,

,

/

/

,

I,,,

1, /

71

I/

77

,,,,!

/,,

!

I

!

1168

1448

1720

aloe

!

Figure 3. Comparison of GC profiles of two types of carrot seed oils. necessary to search the entire library file of 614 entries for each compound in the sample. The identification of the entire sample file containing about 100 compounds is completed in just 3-5 min. In contrast, 1-2 h are usually required if the search is accomplished manually with a commercial MS library. With the aid of retention indexes, isomers can be differentiated where the conventional mass spectral match fails. For example, many monoterpenes have very similar mass spectra, yet with a large spread of retention indexes. The values of a few common monoterpenes generated on a OVlOl fused silica column are cited as follows: tricyclene (918), a-thujene (925), a-pinene (932), a-fenchene (940), camphene (948), /3-phellandrene (1022), limonene (1026), cis-0-ocimene (1032), trans-p-ocimene (1041), and y-terpinene (1045). Clearly, a retention index window of 5 units is sufficient to separate these monoterpenes while a window greater than 7 units will have a deleterious effect on the search result. One probable disadvantage is that each laboratory would have to build its own library file because the reproducibility of retention time and mass ion intensity often vary much depending on experimental conditions, such as column condition and ion source tuning. Library files generated in one laboratory might not be compatible with the experimental conditions set in another laboratory. Comparison between Two Profiles. The usefulness of profile comparison between mixtures in quality control is apparent and straightforward. Samples are routinely compared to a standard by calculating functions F,, or F21. Each sample could be a starting material (e.g., an essential oil) from various sources or a finished product (e.g., a perfume) produced at different times. Low values of F12and Fzlindicate a deviation from reference. A closer examination shall follow to determine the cause of deviation. Three different types of results are shown in the following examples. In Figure 1,geranium oil Bourbon is compared to geranium oil Egyptian. Both F12 and FZ1are rather high, indicating a very similar composition between these two types of geranium oils. In the next example (Figure 2) two cardamon oils purchased from different sources are compared. The differences between them are more obvious, and both F12 and F2, are accordingly lower. In Figure 3, a sample of carrot seed oil is compared to a reference in our library. F12is high, but Fzl is low, indicating that the profile of sample mixture is matched to the reference profile but not vice versa. This means that most substances in the sample mixture are found in the reference, while many other components in the reference

M

6ea

Figure 2. Comparison of GC profiles of two types of cardamon oils.

888

6064

54

ANALYTICAL CHEMISTRY, VOL. 57, NO. 1, JANUARY 1985

Table I. Detection of Geranium Oil in a Perfume Sample: Comparison of the Areas of 15 Index Peaks in Reference and Sample Files

Table 11. Detection of Cedar Leaf Oil in a Perfume Sample: Comparison of the Areas of 15 Index Peaks in Reference and Sample Files

no.

no.

1

2 3 4 5 6 7 8 9 10 11 12 13 14 15

area (ref)

area (sample)

name

1000.00 621.56 298.88 179.53 171.21 101.59 40.17 27.46 26.63 19.01 17.44 10.66 18.56 16.49 7.57

955.37 626.91 346.18 139.34 175.73 83.86 13.12 45.68 22.41 46.19 12.44 19.00

citronellol geraniol citronellyl formate menthone geranyl formate isomenthone geranyl butyrate a-terpineol rose oxide geranyl acetate geranyl tiglate citronellyl acetate citronellyl butyrate geranyl propionate linalyl oxide

0.00 39.20 12.86

1 2 3 4 5 6

7 8 9 10 11 12 13 14

15

area (ref)

area (sample)

name

1000.00 381.35 64.28 51.40 139.21 42.04 13.36 8.84 6.23 5.84 5.69 5.69 4.48 1.98 1.08

488.87 408.70 128.27 133.35 190.60 42.67 236.22

thujone fenchone isobornyl acetate camphor isothujone 4-terpineol a-terpinyl acetate caranyl acetate a-cyclocitrylacetate fenchyl acetate borneol a-terpineol 0-terpineol ethyl isovalerate fenchyl alcohol

0.00 0.00 3.94 26.21 50.08

0.00

0.00 0.00

GERANIlAl

BOURBOH MTRI. UHGERER

I

600

I’

928

1248

1568

I

1888

2200

I

Figure 4. Detection of geranium oil in a perfume sample.

are absent from the sample mixture. In a reversed case, where the sample contains extra components, a low F12and high F2, will result. Detection of Essential Oils in Complex Mixtures. To evaluate our programs in detecting essential oils in complex mixtures, the following two series of experiments were conducted. In the first series, commercial perfumes were spiked with various essential oils a t a 2% level. After the identification of each component in the sample mixture, a search of a library containing 43 essential oils was performed. Each time, the spiking essential oils passed both prefilters and appeared high on the printout of matching results. However, the values of function Fzl showed a wide range of variation from case to case. Two examples are discussed below. The first example illustrates the detection of geranium oil in a perfume. The results are persented in Table I. The 15 index peaks of geranium oil are listed in the column under “name”. Values under “area (ref)” are their relative concentration in the geranium oil reference file. Values under “area (sample)” are their relative concentration found in the sample. Note that the relative concentration of these index compounds in the sample are scaled up such that they are comparable to those in the reference file. The fact that 14 out of 15 index peaks of geranium oil were recognized in the sample mixture proved this detection to be successful. The F,, value calculated based on these index peaks is 836. Figure 4 displays the pattern of the sample mixture (upper trace) and geranium oil

351

IIlll,

I./, .

, .

I

352

ANALYTICAL CHEMISTRY, VOL. 57, NO. 1, JANUARY 1985

results have been presented in another publication in more details (14). Accuracy of Search. For a complex mixture, it is not unusual that certain components are incompletely resolved from each other even on a modern capillary column. Peak separation is, however, essential for a compound to be identified in a computer-automated process. A poor resolution in a GC/MS experiment causes the sample file to be incomplete and/or erroneous. The subsequent comparison with references and calculations of functions F,, and F2, then become meaningless. To remedy this, unresolved components are manually identified by mass spectral search and are amended to sample file by the editing program. It is sometimes necessary to use a large sample size in order to observe trace components. This causes a serious overloading of large components. Overloaded peaks affect the resolution of all nearby peaks and causes their retention times to deviate from the normal values. In these cases, the editing program may be used to correct retention time data in the sample file. Once the retention times are corrected, no further difficulty is encountered in identification. Finally, error can be caused by instability of instruments. The column may have deteriorated to a degree that resolution is lost and retention data become irreproducible. The oven temperature may be unstable, causing further error in retention times. All these experimental factors have to be considered and carefully calibrated in order that an accurate search can be achieved. Efficiency of Search. The manner in which an essential oil is compiled into a library has a profound effect on the efficiency of search. The two most important criteria to be considered in building index tables of an essential oil are the intensity and specificity of the selected index peaks. These peaks should be among the most intense components so that they can be recognized easily in a complex mixture. They should also be relatively specific to a particular oil such that the chance for another oil containing the same index peaks is small. Commonly occurring substances such as monoterpene and sesquiterpene hydrocarbons should not be selected as index peaks. A logical procedure would be to choose the most intense peak in the first step. If this peak is not specific enough, it will be replaced by the next most intense peak. The same procedure is repeated until 10-15 index peaks are selected. The very nature of each essential oil also determines how efficiency it can be identified. Generally speaking, the more specific components an oil contains, the easier it is to be recognized. As far as their chemical composition is concerned, most essential oils can be categorized into one of the following three types: (a) those oils containing many specific components a t significant levels, (b) those oils containing a single major component constituting 70% of more of the total composition, and (c) those oils consisting mostly of terpene hydrocarbons. Examples of class a are geranium, spearmint, and peppermint oils. These oils can usually be easily detected by the presence of their major components. A detection level of 1-5% can be routinely achieved. Oils belonging to class b can be indicated by its single major components a t a 0.05% level. However, since most other components are present at a much lower level, further confirmation is difficult. The calculation

of functions Fl2 and F,, also becomes less accurate. For example, both estragon and basil oils contain estragol as the single major component (88% and 7270, respectively). The differentiation between them would be based on the detection of other trace components. Other examples of class b are anise oil (88% anethole), cinnamon bark oil (96% cinnamic aldehyde), and clove bud oil (87% eugenol). Finally, there are many oils that are composed mostly of terpene hydrocarbons. Terpene hydrocarbons are obviously not useful in differentiating essential oils since they occur in almost all oils. For example, terpene hydrocarbons constitute approximately 95% of the total composition in lemon oil. Its identification then depends on the detection of the small amount of aldehydes present in the oil, e.g., geranial, neral, octanal, nonanal, and decanal. The analysis of this class of oils is least efficient. Overall, a successful identification of an essential oil largely depends on the correct experimental conditions, the very nature of each oil, and the complexity of the sample. A properly compiled library can make the task much more efficient and accurate. The building of a library is still a very time-consuming process. Data of each individual reference must be acquired, analyzed, edited, and added into the library. Good knowledge of the chemistry of essential oils is also required. Nevertheless, with continuous improvement and refinement in the programs, some of the technical difficulties can be removed eventually. It is also our hope that more attention will be devoted to this subject in the future as a result of this publication.

ACKNOWLEDGMENT We thank Axel Kiesslich for this technical assistance and Abraham Jacobowitz and Richard Bertels for providing essential oil samples. Registry No. Citronellol, 106-22-9;geraniol, 106-24-1;citronellyl formate, 105-85-1; menthone, 89-80-5; geranyl formate, 105-86-2;isomenthone, 491-07-6;geranyl butyrate, 106-29-6;aterpineol, 98-55-5;rose oxide, 16409-43-1;geranyl acetate, 105-87-3; geranyl tiglate, 7785-33-3; citronellyl acetate, 150-84-5;geranyl propionate, 105-90-8;linalyl oxide, 5989-33-3;thujone, 546-80-5; fenchone, 1195-79-5;isobornyl acetate, 125-12-2;camphor, 76-22-2; isothujone, 471-15-8; 4-terpineol, 562-74-3; a-terpinyl acetate, 80-26-2; fenchyl acetate, 13851-11-1;borneol, 507-70-0.

LITERATURE CITED Ziemer, J. N.; Perone, S. P.; Caprioii, R. M.; Seifert, W. E. Anal. Chem. 1979, 5 1 , 1732. Gates, s. C.; Smisko, M. J.; Ashendel, C. L.;Young, N. D.; Holland, J. F.; Sweeley, C. C. Anal. Chem. 1978, 5 0 , 433. Rephaeiian, L. A. "Abstracts of Papers", 181st National Meeting of the American Chemical Society, Atlanta, GA, 1981; American Chemical Society: Washington, DC, 1981;FUEL 33. Lea, R. E.; Bramston-Cook, R. Anal. Chem. 1983, 5 5 , 626. Demirgian, J. C. J . Chromatogr. Sci. 1984, 22, 153. Russel, G. F. I n "Chemistry of Foods and Beverages: Recent Developments"; Charalambous, G., Ingiett, G., Eds.; Academic Press: New York, 1082; p 129. Martisen, D. P. Appl. Spectrosc. $981, 3 5 , 255. Hertz, H. S.; Hites, R. A.; Biemann, K. Anal. Chem. 1971, 4 3 , 681 Rasmussen, G.T.; Isenhour, T. L. J , Chem. I n f . Comput. Sci. 1979, 19. 179. Kowalskl, 8 . R.; Bender, C. F. Anal. Chem. 1972, 4 4 , 1405. Justlce, J. B.; Isenhour, T. L. Anal. Chem. 1974, 48, 223. Kowalski, B. R.; Bender, C. F. J . Am. Chem. SOC. 1972, 9 4 , 5632. Kowalskl, B. R.: Bender, C. F. J . Am. Chem. SOC. 1973, 95, 686. Chien. M. Perfum. Flavor. 1984, 9 , 167.

RECEIVED for review June 4,1984. Accepted October 16,1984.