ARTICLE IN PRESS

0958-6946/$ - se

doi:10.1016/j.id

�CorrespondE-mail addr

International Dairy Journal 17 (2007) 226–234

www.elsevier.com/locate/idairyj

Assessment of Trentingrana cheese ageing by proton transferreaction-mass spectrometry and chemometrics

Eugenio Apreaa,b,�, Franco Biasiolia, Flavia Gasperia, Daniela Motta,Federico Marinic, Tilmann D. Markb,d

aIASMA Research Center, Agrifood Quality Department, Via E. Mach, 1, 38010 San Michele all’Adige, TN, ItalybInstitut fur Ionenphysik, Universitat Innsbruck, Technikerstr. 25, A-6020 Innsbruck, Austria

cDipartimento di Chimica, Universita di Roma ‘‘La Sapienza’’, P.le A. Moro 5, 00185 Roma, ItalydDepartment of Plasmaphysics, University of Bratislava, SK-84248 Bratislava, Slovak Republic

Received 23 March 2005; accepted 20 February 2006

Abstract

Proton transfer reaction-mass spectrometry (PTR-MS) data have been analysed by chemometric techniques to monitor cheese ageing

by means of on-line direct head-space gas analysis. Twenty cheese loaves of Trentingrana, a trademarked cheese produced in northern

Italy, of different origin and ripening degree, were sampled over the whole Trentingrana production area. An increase of the spectral

intensity with ripening has been observed for most of the PTR-MS peaks and a univariate analysis identified 16 mass peaks that were

significantly different for ripened and young cheeses, respectively. Moreover, the usefulness of different discriminant analyses and class

modelling techniques have been investigated. Discriminant Partial Least Squares analysis, while indicating average behaviour and

possible outliers, was not able to correctly classify all samples. Soft class modelling performed better and allowed a 100% correct

classification. Partial least square calibration predicted the ageing time of each loaf with reasonable accuracy with a maximum cross-

validation error of 3.5 months.

r 2006 Elsevier Ltd. All rights reserved.

Keywords: Grana cheese; Cheese ripening; Proton transfer reaction-mass spectrometry; Head-space analysis; Chemometry; Volatile organic compounds

1. Introduction

Grana Padano is one of the best known Italian hardcheeses and its long tradition and typicality is wellrecognised and certified by a protect denomination oforigin (EC, 1996). Within the consortium of GranaPadano, a more specific geographic identification with itsown trademark is recognised as ‘‘Trentingrana’’ (TG)referring to loaves produced in the territory of Trentoprovince (northern Italy). TG, belonging to the family ofGrana Padano, is produced with partly skimmed, rawcows’ milk acidified with a mixture of Lactobacillus andcoagulated by the addition of bovine liquid rennet but,having a more restrictive production protocol, the useof lysozyme is not allowed. Further difference is the

e front matter r 2006 Elsevier Ltd. All rights reserved.

airyj.2006.02.008

ing author. Tel.: +390461 615388; fax: +39 0461 650956.

ess: eugenio.aprea@iasma.it (E. Aprea).

prohibition to use silage for the cows’ feeding (Carini &Lodi, 1997). Even if the production of TG is relativelysmall (about 4000 ton year�1; consortia personal commu-nication) when compared with that of the direct competi-tors (about 125 000 ton year�1 for Grana Padano; andabout 113 000 ton year�1 for Parmigiano; http://www.gra-napadano.com, www.crpa.it), it is relevant for the localeconomy.Cheese flavour compounds result from the action of

microorganisms and enzymes on the carbohydrates, fatsand proteins of the milk and curd and the principalbiochemical pathways for their formation have beenreported by many authors (Kristoffersen, 1973; Manning,1979; Engels & Visser, 1994; Fox, Singh, & McSweeney,1995; Fox & Wallace, 1997; McSweeney, 1997) andreviewed by McSweeney and Sousa (2000). The volatilecomponents of dairy products have been widely studied(Aston & Dulley, 1982; Law, 1984; Manning, Ridout, &

ARTICLE IN PRESSE. Aprea et al. / International Dairy Journal 17 (2007) 226–234 227

Price, 1984; Yvon & Rijnen, 2001; Adda, 1986; Liu,Holland, & Crow, 2004) and more than 600 volatilecompounds have been identified in cheese (Maarse &Visscher, 1989; Stratton, Hutkins, & Taylor, 1991). Inparticular, the volatile organic compounds (VOCs) presentin Parmigiano Reggiano and Grana Padano have beenextensively studied by chromatographic techniques(Bottazzi & Battistotti, 1974; Dumont, Roger, & Adda,1974; Meinhart & Schreier, 1986; Barbieri et al., 1994;Moio & Addeo, 1998). An alternative approach to theexpensive and time-consuming chromatographic techni-ques is the use of direct headspace-mass spectrometry foranalysis of VOCs (Marcos Lorenzo, Perez Pavon, FernandezLaespada, Garcya Pinto, & Moreno Corsero, 2002).

Along this line we characterised TG cheese by means of adirect headspace analysis using proton transfer reaction-mass spectrometry (PTR-MS). As indicated by Boscaini,Van Ruth, Biasioli, Gasperi, and Mark (2003) whenanalysing similar cheese samples, the choice of PTR-MShas several advantages because of a number of interestingcharacteristics: (i) PTR-MS is fast, time-dependent varia-tions of the headspace can be monitored with a sub-seconds time-resolution and the headspace of a sample canbe measured in a few minutes; (ii) samples are not subjectedto any previous treatments thus reducing the risk ofartefacts; (iii) mass spectral intensity can be transformedinto absolute headspace concentrations in principle with-out any calibration with external standards; (iv) PTR-MShas a detection limit in the ppt range (Lindinger, Hansel, &Jordan, 1998) and a high dynamic range of several ordersof magnitude. Probably, the main drawback of a directPTR-MS analysis is that, lacking of a chromatographicseparation and of an exhaustive database on purecompounds fragmentation, the interpretation of PTR-MSspectra in the case of complex mixtures, remains a difficultissue. However, for the majority of the compoundsreported in the paper the fragmentation pattern acquiredin standard condition (the same we used) can be found in

Table 1

The factories of the Trentingrana consortium where the analysed cheeses have

Cheese making factory code Age of the loaves (months) and production p

Young Ripened

Ab (9) Aug‘01 (21)

Bb (11) Sep‘01 (22)

C (8) Sep‘01 (25)

D (11) Jul‘01 (22)

E (12) Jul‘01 (23)

F (11) Jul‘01 (28)

G (14) May‘01 (23)

H (11) Aug‘01 (23)

I (9) Aug‘01 (19)

J (12) Jul‘01 (18)

aT ¼ traditional management system: small farms (o20 cows), milk pro

concentrates. M ¼ modern management system: medium (20–40 cows) or large

high level of concentrates.bA and B refers to the same cheese factory that provided four loaves in two

Buhr, van Ruth, and Delahunty (2002) and a furtherinterpretation can be based on the work of Boscaini et al.(2003) on similar samples. Tentative peak attribution hasbeen assisted by peaks correlation, isotopic ratios, and byfurther comparison between headspace evolution of cheeseand pure compounds. The possibility to follow on line thedecrease of compounds in the headspace of liquid solutionsallows sometimes the separation of isobars (Karl, Yer-etzian, Jordan, & Lindinger, 2003; Pollien, Jordan,Lindinger, & Yeretzian, 2003; Yeretzian, Jordan, &Lindinger, 2003). This has been used, e.g., to estimateacetaldehyde in cheese based on the intensity of the signalat m=z ¼ 45 (Aprea et al., 2003).We studied Trentingrana samples of different ages from

different origins in order to demonstrate the possibility tocharacterize certain features of TG cheese, i.e., in particularfor age prediction by PTR-MS analysis. Different chemo-metric methods have been applied to the PTR-MSfingerprint data and their usefulness for age predictionhas been evaluated. In a recent work (Biasioli et al., 2006)correlation among PTR-MS spectral fingerprinting andodour and flavour sensory profile of TG has beeninvestigated.

2. Materials and methods

2.1. Cheese samples

Twenty cheese loaves of TG, aged between 8 and 28months, were obtained from the main cheese-factorieslocated in Trentino province (Italy) and certified by the‘‘Trentingrana’’ consortium (the actual sampling took aperiod of 3 months). Each cheese factory provided twosamples: a ‘‘ripened’’ one (not less than 18 months) and a‘‘young’’ one (less than 15 months) (Table 1). All the cheesefactories followed the same protocol (Battistotti &Corradini, 1993) for the cheese production, the onlyrelevant declared difference was that for two factories

been sampled

eriod Production: loaves year�1 Management systema

Aug‘00 6292 M,T

Oct‘00 6292 M,T

May‘00 5226 T

Aug‘00 5260 M,T

Aug‘00 6480 M,T

Mar‘00 6622 T

Aug‘00 9182 T

Aug‘00 3218 T

Oct‘00 8750 M

Jan‘01 6745 M

duction o25L day�1cow�1, feeding with local hay and low level of

(440 cows) farms, milk production of 25–40 L day�1cow�1, feeding with

different sampling.

ARTICLE IN PRESSE. Aprea et al. / International Dairy Journal 17 (2007) 226–234228

(E, J) the milk was cooled down and kept at 12–14 1Covernight the day before the cheese making.

From a piece of cheese of about 5 kg (height10.5–11.5 cm, radius 20–24 cm), corresponding to about a1/8 of a loaf we removed the outer layer (at least 4 cm). Thesample for the analysis was a slice of about 2.5� 4� 12 cmfrom the remaining inner part of the loaf. This slice wasgrated and well mixed to reduce the variability of thesamples due to the differences in the different parts of thecheese loaf. An amount of 2.5 g of this grated cheese wasdeposited in glass bottles of 120mL (Supelco, Bellefonte,USA) sealed with a cap provided with teflon/siliconeseptum and stored in a refrigerator at 4 1C for no longerthan 7 h without any further treatment. After the bottlesclosure they were not opened and the headspace wassampled through a needle piercing the septum. For eachcheese three vials were prepared.

2.2. PTR-MS

The PTR-MS technique has been extensively discussed ina series of review papers (Hansel et al., 1995; Lindingeret al., 1998), here we will just recall some general aspects.Introduced by Lindinger and co-worker in 1993 (Lindin-ger, Hirber, & Paretzke, 1993; Lagg, Taucher, Hansel, &Lindinger, 1994), it is based on a novel design of thechemical ionisation method (Munson & Field, 1966). Thesample gas is continuously introduced into a drift tubewhere it is mixed with H3O

+ ions formed in a hollowcathode ion source. Volatile compounds that have protonaffinities higher than water (proton affinity of H2O:166.5 kcalmol�1) are ionised by proton transfer fromH3O

+, mass analysed in a quadrupole mass spectrometerand eventually detected as ion counts per second (cps) by asecondary electron multiplier. A commercial PTR-MSapparatus (Ionicon GmbH, Innsbruck, Austria) was usedand the mass spectrometric data were collected over a massrange from m/z 20 to m/z 259 using a dwell time of 0.2 s permass (in each cycle a complete mass spectrum up to mass259 amu is monitored within a time span of 48 s) underdrift tube condition of 120–130 Td (Td ¼ Townsend; 1Td ¼ 10�17 V cm2mol�1). Each sample was measured foreight cycles and the mean of cycles 3–7 was used for furtheranalysis. Consecutive samples were alternated with blank(empty bottle) whose signal was subtracted from samplespectra. The data collected were converted in ppb accord-ing to a procedure given by Lindinger et al. (1998). Weassumed the reaction rate constant to be 2� 10�9 cm3 s�1

for all compounds, thus introducing a systematic error thatis often reasonable (Lindinger et al., 1998). This effect mustbe considered if comparing our data with results of othertechniques but it does not affect the proposed chemometricanalyses which require only that the measuring conditionsare constant. The problem of signal drifts related to thefingerprinting based-sensor techniques (Kress-Rogers,1996) should be overcome because the conversion fromion counts into the absolute concentration (ppb) takes into

account the parameters that can change in time and withdifferent apparatuses (Lindinger et al., 1998). Moreover wemeasured mostly young and ripened cheeses together(randomising the order of samples and replicates) thus aninstrumental shift, if present, cannot induce a separationbetween the investigated classes of cheese.

2.3. Headspace analysis

Samples were placed in a water bath at 36.6 1C one hourbefore and during PTR-MS measurements. The vial wasconnected via a heated (70 1C) capillary line made ofuncoated deactivated fused silica tubing with an innerdiameter of 0.25mm (Supelco, Bellefonte, USA) terminat-ing in a stainless steel needle to the drift tube of the PTR-MS (Biasioli, Gasperi, Aprea, Colato et al., 2003). Thedifference in pressure between the drift chamber(2.0� 10�3 bar) and the vial (atmospheric pressure) gen-erates a flux flow of about 10 cm3min�1 (due to the lineimpedance), this flux correspond to a velocity of about3.4m s�1 resulting in a time of about 0.6 s for the gas tocover the entire length of the transfer line. To avoid apressure drop during the headspace sampling, the removedgas mixture was replaced with pure nitrogen gas (SOLs.p.a., Italy; purity: 99.999%) through a second stainlesssteel needle connected to a nitrogen cylinder.The samples were measured with a time interval of

20min and considering 1 h of conditioning in the waterbath, 5–7 h passed between the measure of the first sampleand the last one. The measuring order of samples andreplicates was randomised to avoid memory effect andnitrogen was flushed, for about 10min, to clean the transferline between two consecutive measurements.

2.4. Data analysis

PTR-MS is a spectrometric technique, whose output is ahigh dimensional vector (hundreds of masses) of inter-correlated data (the absence of pre-separation producesinterference of different compounds or fragments on thesame spectrometric peaks). The analysis of this kind ofdata is efficiently performed by multivariate analysis in twosteps: (i) data compression to reduce the dimensionality ofthe problem without a significant loss of information and(ii) classification to identify possible interesting structure inthe data set.In a first approach to the problem, data compression was

performed by means of discriminant Partial Least Squareanalysis (dPLS) (Kemsley, 1998) that is a restriction of thePartial Least Square technique (Beebe & Kowalsky, 1987),where the additional information (category index) is (i)used to maximise between-class separation and is (ii) usefulin finding the variates responsible for group differences.Linear discriminant analysis (LDA) (Duda, Hart, & Stork,2001) is applied after this compression phase. The LDAmethod is a pure classification technique, it uses linearboundaries (hyperplanes) to delimit the class regions.

ARTICLE IN PRESSE. Aprea et al. / International Dairy Journal 17 (2007) 226–234 229

In this way, the classification rule can be used forpredictions of unknown samples. However, to have anaccurate estimation of the predictive ability of the model inthe case of unknown samples, a further validation step isneeded. In this study, a ‘‘leave-one-out’’ cross-validationapproach was used (Stone, 1974; Good, 1999). It should bepointed out that the predictive ability of a classifier couldalso be affected by data pre-treatment. In particular wenoted in previous studies that LDA on PTR-MS spectraperforms better if each spectrum is normalised to unit areabefore any statistical processing (Biasioli, Gasperi, Aprea,Colato et al., 2003; Biasioli, Gasperi, Aprea, Mott et al.,2003). Therefore also in this study the spectra werenormalized before data analysis.

A different approach to the problem of pattern recogni-tion is the use of class-modelling instead of the use of pureclassification techniques. Here we used the UNEQ (UN-EQual disperses classes) technique which focuses on onecategory at a time, defining a boundary to separate aspecific class from the rest of the universe. The class spaceis built according to the Mahalanobis distance as theconfidence hyperellipsoid that accounts for 95% of theprobability in the case of multivariate normal distribution.This approach seems to be more powerful and moreappropriate when dealing with real-world problems.However, its implementation requires a high ratio betweenthe number of samples for each class and the number ofexperimental variables (more than 3), so a preliminaryvariable reduction stage is often necessary. In this study thevariable reduction has been carried out using a stepwiseprocedure, based on the value of Wilks’ lambda, which isan inverse multivariate measure of the discriminatingability of a set of variables (Wold et al., 1984).

The preliminary data exploration, standard univariatedata analysis and analysis of variance (two groups: youngand ripened, ten samples per group, three replicates foreach sample) were performed with the software packageStatistica 5.0 (StatSoft, Inc., Tulsa, OK, USA) on spectranormalised to unit area.

Discriminant PLS analysis, as reported by Kemsley(1998) and already tested on PTR-MS spectra (Biasioli,Gasperi, Aprea, Colato et al., 2003) was implemented bymeans of the software Win-DAS (John Wiley & Sons, Ltd:Chichester, UK, 1998). All the other computations havebeen performed using the package V-Parvus 2003 (Forina,Lanteri, Armanino, Cerrato-Oliveiros, & Casolino, 2003).

3. Results and discussion

3.1. Descriptive statistics

The original data matrix consisted of 60 objects (rows),corresponding to the cheese samples, and 240 variables(columns), corresponding to the headspace concentrationof all the measured masses from m/z 20 to m/z 259, datawere converted in ppb. We expected a great variabilityamong the samples, because, even if controlled by a single

consortium, GT cheese is produced, ripened and stored insmall cheese factories scattered in the Trentino provincevalleys (200ma.s.l.) and mountains (up to 1000ma.s.l.).At first glance of the data seems that the spectra of the

two groups (‘‘young’’ and ‘‘ripened’’) are very similar butthe average intensities, anyway, even if an overall increaseof the intensity is the main effect, every sample has its ownspectrum with peculiar details. Despite this significantdifference between the two groups, on the contrary we didnot observe significant differences among replicates noreffects related to the measuring session.From an ANOVA analysis several masses can be

identified which are significantly different for the twogroups, i.e., 16 masses with a confidence greater than99.9% (m=z ¼ 37, 43, 45, 47, 61, 63, 65, 71, 89, 91, 103,107, 109, 115, 117, 119, 145), 24 with a confidence between99.0% and 99.9%, and 24 with a confidence between 90%and 99%. Here we do not count the peaks which intensityis that expected for isotopes of other masses. We canconclude that the two groups are statistically different andthat many peaks have a significantly higher level in the caseof ripened cheeses. For the most significant masses atentative identification based on literature data (Boscainiet al., 2003; Bottazzi & Battistotti, 1974; Dumont et al.,1974; Meinhart & Schreier, 1986; Barbieri et al., 1994;Moio & Addeo, 1998) and pattern fragmentation of purestandards (Buhr et al., 2002) is given in Table 2.In general, an increase of the peak intensity for the

ripened cheeses is evident. Moreover, it seems that in youngsamples the headspace composition is more homogenous,while during the ripening the sample of each cheese factoryfollows a different path and thus differences betweensamples are increasing. For example, the signals of themasses related to esters (m/z 89, 103, 117, 145) increasewith ripening for the majority of the samples. Esters arerecognised as common constituents of cheese flavour.Esterification reactions occur between short-chain andmedium-chain fatty acids and primary and secondaryalcohols derived from lactose fermentation or from aminoacid catabolism during cheese ripening (Curioni & Bosset,2002). High concentration of ethyl esters in ‘‘Parmesan’’cheese (Dumont et al., 1974; Meinhart & Schreier, 1986)and in ‘‘Grana Padano’’ cheese (Moio & Addeo, 1998) hasbeen reported previously. In particular, ethyl hexanoate isthe major ester present in both cheeses followed inabundance by the ethyl esters of butanoic, octanoic anddecanoic acids (Meinhart & Schreier, 1986; Moio & Addeo,1998). The most intense peaks for these compounds areexpected at m/z 145 for ethyl hexanoate, m/z 89 and 117 forethyl butanoate, m/z 173 for ethyl octanoate and m/z 201for ethyl decanoate (Buhr et al., 2002).As an example for the tentative identification of a PTR-

MS peak we report the case of mass 45. The most intensesignal for ripened cheeses was recorded at m/z 45 with anaverage value of 17.5 ppmv (CV% 92) showing an elevatedbroadening of the sample, while in young cheeses a muchlower average value of about 1.0 ppmv (CV% 61) was

ARTICLE IN PRESS

Table 2

Masses, with tentative identification, showing significant different

intensities (po0.001, on data normalised to unit area) in young and

ripened cheesesa

m/z Class All samples Tentative identificationb

Mean

(ppbv)

CV (%)

43 Young 2376 35 Fragment common to

several compounds

Ripened 3898 43

45 Young 1038 63 Acetaldehyde

Ripened 17503 94

47 Young 1152 64 Ethanol

Ripened 2255 46

61 Young 1197 45 Acetic acid; acetyl esters

fragment

Ripened 1759 43

63 Young 68 40 Acetaldehyde-water

clusterc

Ripened 913 95

65 Young 230 56 Ethanol-water cluster

Ripened 449 49

71 Young 120 46 C-5 alcohols; C-4 acids

fragment

Ripened 299 66

89 Young 379 47 Ethyl acetate; butanoic

acid

Ripened 1120 75

91 Young 13 21 –

Ripened 32 49

103 Young 18 52 Methyl butanoate;

isovaleric acid

Ripened 36 53

107 Young 36 46 Benzaldehyde; ethyl

benzene; o,p,m-xylene

Ripened 109 77

109 Young 6 30 Trans-2-octenal; d-carvone fragment

Ripened 7 36

115 Young 52 50 2-heptanone

Ripened 95 51

117 Young 39 57 Hexanoic acid; C-6 esters

Ripened 240 80

119 Young 4 44 2-butoxyethanol

Ripened 7 46

145 Young 4 124 Ethyl hexanoate

Ripened 29 122

aHeadspace average concentration (three replicates) expressed in ppbV

and coefficient of variation (%) is reported for the selected masses.bIdentification made comparing literature data (VCF database) and

fragmentation patterns (Buhr et al., 2002).cFrom the consideration reported in the text and from the data about

GC-O reported in Boscaini et al. (2003) the presence of dimethyl sulfide

for GT is below the instrumental error.

E. Aprea et al. / International Dairy Journal 17 (2007) 226–234230

observed. We attributed the signal recorded at m/z 45almost exclusively to the presence of protonated acetalde-hyde (acetaldehyde molecular weight 44 amu). We esti-mated the possible contribution to the signal recorded atm/z 45 from fragments of other compounds comparing thespectra of cheese with the fragmentation pattern of purecompounds. For example, the contribution of 3-methyl-

butanal at mass 45 should be negligible. Mass 69 is themain fragment of 3-methyl-butanal (about 73%), if wesuppose that mass 69 is only 3-methyl-butanal (is not thisthe case), its contribution to the intensity recorded at m/z45 should be below 3%. In addition to the fermentationoperated by microorganisms (lactose metabolism) (Ott,Germond, & Chaintreau, 2000), acetaldehyde can alsoform in threonine degradation, a process that could be ofimportance during cheese ripening (Engels, Dekker,de Jong, Neeter, & Visser, 1997). An increase of itsconcentration during ripening was reported for SwissEmmental (Bosset, Buetikofer, Gauch, & Sieber, 1997)and Pecorino Sardo (Larrayoz, Addis, Gauch, & Bosset,2001). Moreover, Boscaini et al. (2003) found that the peakat m/z 45 is the most intense in TG headspace and that,being about five and 15 times higher than in ParmigianoReggiano and Grana Padano, respectively, it could be usedas a marker for the discrimination among these cheeses. Inthis study, having a wider number of samples, we confirmthat for ripened GT the signal at m/z 45 is the predominantin the headspace spectra and furthermore it increases withripening. From the headspace intensity of signal at m/z 45(ppb), we attributed to acetaldehyde, we estimated the mgof acetaldehyde for kg of cheese (�3mg kg�1; Aprea et al.,2003) and, in spite of headspace concentration seems veryhigh, the results are compatible with those reported forParmesan cheese (3.4–7.6mg kg�1; Barbieri et al., 1994).

3.2. Classification and class-modelling

The present PTR-MS data matrix (60 samples� 240masses) of the analysed samples was then used to build aclassification model to discriminate between ‘‘young’’ and‘‘ripened’’ cheeses. At first, discriminant partial leastsquares (dPLS) analysis was carried out, considering eachreplicate as an individual sample. The first two dPLS scoresare plotted in Fig. 1, where the separation between the twoclasses is evident. Furthermore, the group of ripenedcheeses, appears to be split into at least two separate sub-classes, one of these groups is formed by the samples HR,ER, JR. The headspace profile of these three samplespresents a lower intensity more similar to the spectra ofunripe cheeses. The cheese-factories E and J use a mixtureof milks from two consecutive milking, in which the milkfrom the evening milking is kept overnight at 12–14 1C.This could partly be at the origin of the observeddifferences. Furthermore, the sample JR has only 18months of ripening so it is reasonable that the headspaceprofile is closer to the young cheeses.Moreover, Linear Discriminant Analysis based on

Mahalanobis distances on the first three PLS scores wasapplied to the present data. In the modelling phase, all thesamples were assigned to the correct group. Then, due tothe small number of samples, a ‘‘leave one out’’ crossvalidation approach was carried out to validate the model.All the validation samples have been correctly classified.Significantly worse results in terms of the overall predictive

ARTICLE IN PRESS

Fig. 2. First loading of dPLS analysis. The underlined masses are those

related with esters.

8

6

4

2

0

-2

-4

-6

-8

-10

-12-10 -8 -6 -4 -2 0 2 4 6

PLS Score 1 (9.7%)

PLS

Sco

re 2

(7.

8%)

ER

HR

JR

Fig. 1. Plot of the first two dPLS scores for all samples. The two scores

separate well the young cheeses (grey symbols) from the ripened (black

symbols). For the ripened cheeses there are at least two groups: samples

JR, HR, ER form a separate group (the three replicates are encircled).

Dis

tanc

e to

Rip

ened

Distance to Young

95%

Fig. 3. Coomans plot for the UNEQ model: young cheeses are

represented as filled circles, while ripened cheeses are shown as filled

triangles.

E. Aprea et al. / International Dairy Journal 17 (2007) 226–234 231

ability are obtained, if a different validation procedure isfollowed. In fact, if a 20-fold cross-validation approach,where the three replicates for each samples are kept out at atime, is adopted, a minor success rate of 85% is obtainedfor the validation set. This indicates that the modelperforms better when it is built using the maximumvariability and also that the misclassification is probablydue to the peculiarities of the misclassified samples. Indeed,the nine samples wrongly assigned are the three replicatesof HR, ER, and JR, the same that form an own cluster indPLS plot (Fig. 1).

In addition to allowing quality control and productclassification, multivariate approaches give also usefulhints for chemical analysis. In fact, an inspection of theloadings for the first latent variables can immediatelyprovide an indication of the masses responsible for theobserved separation between the classes; it is then possibleto try to correlate these masses with chemical compounds.In particular, the analysis of the first loading (Fig. 2), thatresembles the shape of PTR-MS data, reveals that theseparation along the first PLS component is mainly due toa group of masses related to esters (m=z ¼ 89, 103, 117,131, 145 and 173). This result confirms our previousobservation, reported above, where, as far as the massesrelated to esters were concerned, higher signals wereobserved for the ‘‘ripened’’ cheeses with respect to thosefor the ‘‘younger’’ ones. In this way, it is possible to explainthe role of a single compound or a group of compounds inthe discrimination.

For the alternative classification method considered, theclass-modelling technique UNEQ, two variables aresufficient to build the optimal UNEQ model (Fig. 3).

As far as the classification and prediction results areinvolved, a 100% rate of correct assignments has beenobtained, even when a 20-fold cross-validation (as the one

reported above for dPLS, with all the replicates joined intothe same cancellation group) was used. So, the class-modelling approach results in a clear improvement of theprediction results. Additionally, the power of class-model-ling techniques is that they can provide the experimenterwith information about how properly each sample isdescribed by any of the class models. This information canbe encoded into two figures of merit, sensitivity andspecificity, defined as the non-error rate for each class andthe percentage of objects of other classes rejected by theclass-model under study respectively. In the present study,while the specificity was 100%, the sensitivity was 97% asindeed one sample for each class (one replicate of JY andone of HR) fell outside the respective 95% class boundary,thus these points being consequently rejected by the classmodel. The same result can also visualized graphically inthe form of a Coomans plot (Fig. 3), whose axes are the

ARTICLE IN PRESS

30

25

20

15

10

5

00 5 10 15 20 25 30

Ageing (months)

Pre

dict

ed c

ross

-val

idat

ed a

gein

g (m

onth

s)

Fig. 4. Predicted cross-validated estimation of cheese ageing vs actual

values.

E. Aprea et al. / International Dairy Journal 17 (2007) 226–234232

distances from each of the two class models, as computedby UNEQ. The figure shows the clear separation betweenthe two classes (all the samples from one class have asignificant distance to the other class model), the twooutlying observations are represented as points outsidethe straight lines designating the corresponding classboundaries.

3.3. PLS modelling of the cheese ageing

A direct modelling of the relation between the mass-spectral fingerprint and the age of the samples wasattempted using a PLS1 approach, coupled to variablereduction based on the relative importance in prediction.The latter procedure is based on the iterative elimination ofthe variable which contribute less to the model, and hasbeen described elsewhere (Eriksson, Johansson, Kettaneh-Wold, & Wold, 1999). A direct PLS1 analysis on all datawithout any conditioning or variable selection provide areasonable good estimation of the age of the loaves with aRMSECV of 3 months with only four latent variables. Amore accurate analysis can be done as follow. The optimalPLS model was then built using 33 experimental variables(m=z ¼ 37, 41, 43, 45, 46, 47, 49, 55, 57, 59, 63, 64, 65, 69,71, 73, 74, 75, 79, 81, 83, 87, 90, 91, 95, 99, 101, 115, 117,121, 135, 139, 145) which appeared to be most correlated tothe response. The choice of the model complexity wasbased on the minimum error in prediction as evaluated byleave-one-out cross-validation, and resulted in the inclu-sion of 23 latent variables, leading to R2 ¼ 99:8% andQ2 ¼ 99:3%. These results correspond to an average errorof less than 0.2 months on the calibration samples and ofabout 0.4 months on the validation set with a maximumabsolute error of 3.5 months (25%). These results can beseen in Fig. 4, where the predicted cross validated y valuesare compared to the actual values. It is worth mentioningthat even if the model with 23 variables is the best, usingonly five LVs it is also possible to obtain a rather goodmodel (R2 ¼ 92:6 and Q2 ¼ 91:4): this overall difference isreflected in the decrease in the minimum and maximumerrors on increasing the number of LVs (from 1.5 and 4.5,respectively, with five LVs to 0.4 and 3.5 with 23) that iscompatible with the analysis performed without anypreliminary data selection.

4. Conclusion

In this paper chemometric procedures implemented onhead-space PTR-MS data were developed to characterizethe ageing of TG cheese from rapid headspace measure-ments and to set models for ageing prediction. In particulartwo procedures were implemented, checked and compared.The first, based on the dPLS procedure for data compres-sion followed by the construction of a LDA model, showedsatisfactory results for the classification and has theadvantage to be more conservative for data interpretation(loadings of dPLS shows the masses responsible for

differences). The second procedure, based on the UNEQmodel showed better performances for the classification,and has the potential advantage that additional groups canbe added to the overall scheme at any time, whereas inLDA a completely new model is required if the number ofgroups changes.The main advantages of using PTR-MS in this kind of

study lies in the minimal treatment of the samples (justgrinding) thus avoiding possible artefacts, the short time ofthe analysis and the low cost of a single measurement.Besides classification/discrimination a series of additionalchemical information are available.The results of this work indicate that ageing of TG can

be monitored, both qualitatively (building a classificationmodel able to discriminate between young and ripenedcheeses) and quantitatively (by means of a PLS regressionmodel, with a low prediction error). These results areparticularly important, as, in addition to the absence ofdefects, this variable is one of the main factors defining theeconomic value of this cheese. Extension of this work to‘‘Grana Padano’’ and ‘‘Parmigiano Reggiano’’ is plannedaiming at finding a rapid tool to characterise and controlthe production of these important products of the Italianfood culture.

Acknowledgements

Work partially supported by the PAT-CNR projectAGRIIND and by the FWF, Wien, Austria.

References

Adda, J. (1986). Flavour of diary products. In G. G. Birch, & M. G.

Lindley (Eds.), Development in food flavours (pp. 151–172). London,

UK: Elsevier.

Aprea, E., Biasioli, F., Gasperi, F., Mott, D., Boscaini, E., Mayr, D., et al.

(2003). Quantificazione di composti volatili in matrici complesse da

misure di spazio di testa: acetaldeide in formaggio. In S. Porretta (Ed.),

ARTICLE IN PRESSE. Aprea et al. / International Dairy Journal 17 (2007) 226–234 233

Ricerca e innovazione nell’industria alimentare, Vol. 6 (pp. 1098–1103).

Pinerolo, Italy: Chiriotti.

Aston, J. W., & Dulley, J. R. (1982). Cheddar cheese flavour. Australian

Journal of Dairy Technology, 37, 59–64.

Barbieri, G., Bolzoni, L., Careri, M., Mangia, A., Parolari, G., Spagnoli,

S., et al. (1994). Study of the volatile fraction of Parmesan cheese.

Journal of the Science of Food and Agriculture, 42, 1170–1176.

Battistotti, B., & Corradini, C. (1993). Cheese, chemistry, physics and

microbiology. In P.F. Fox (Ed.), Italian cheese (2nd ed.) (pp. 221–233).

London, UK: Chapman & Hall.

Beebe, K. R., & Kowalsky, B. R. (1987). An introduction to multivariate

calibration and analysis. Analytical Chemistry, 59, 1007A–1017A.

Biasioli, F., Gasperi, F., Aprea, E., Colato, L., Boscaini, E., & Mark, T.

(2003). Fingerprinting mass spectrometry by PTR-MS: Heat treatment

vs. pressure treatments of red orange juice—A case study. International

Journal of Mass Spectrometry, 223 & 224, 343–353.

Biasioli, F., Gasperi, F., Aprea, E., Endrizzi, I., Framondino, V., Marini,

F., et al. (2006). Correlation of PTR-MS spectral fingerprints with

sensory characterisation of flavour and odour profile of ‘‘Trentin-

grana’’ cheese. Food Quality and Preferences, 17(1&2), 63–75.

Biasioli, F., Gasperi, F., Aprea, E., Mott, D., Boscaini, E., Mayr, D., et al.

(2003). Coupling proton transfer reaction-mass spectrometry with

linear discriminant analysis: A case study. Journal of Agricultural and

Food Chemistry, 51, 7227–7233.

Boscaini, E., Van Ruth, S., Biasioli, F., Gasperi, F., & Mark, T. D. (2003).

Gas chromatography-olfactometry (GC-O) and proton transfer

reaction-mass spectrometry (PTR-MS) analysis of the flavor profile

of grana padano, parmigiano, reggiano and grana trentino cheeses.

Journal of Agricultural and Food Chemistry, 51, 1782–1790.

Bosset, J. O., Buetikofer, U., Gauch, R., & Sieber, R. (1997).

Reifungsverlauf von in Folien verpacktem Emmentaler Kase, mit

und ohne Zusatz von Lactobacillus casei subsp. casei II. Gaschroma-

tographische Untersuchung einiger fluchtiger, neutraler Verbindungen

mit Hilfe einer dynamischen Dampfraumanalyse. Lebensmittel-Wis-

senschaft & Technologie, 30, 464–479.

Bottazzi, V., & Battistotti, B. (1974). Head space chromatograms of

carbonyl compounds in the volatile components of Grana cheese.

Scienza e Tecnica Lattiero-Casearia, 25, 11–19.

Buhr, K., van Ruth, S., & Delahunty, C. (2002). Analysis of volatile

flavour compounds by proton transfer reaction-mass spectrometry:

Fragmentation patterns and discrimination between isobaric and

isomeric compounds. International Journal of Mass Spectrometry,

221(1), 1–7.

Carini, S., & Lodi, R. (1997). Grana Padano. Consideration of starter

culture and lysozyme. Latte, 22, 76–79.

Curioni, P. M. G., & Bosset, J. O. (2002). Key odorants in various cheese

types as determined by gas chromatography-olfactometry. Interna-

tional Dairy Journal, 12, 959–984.

Duda, R. O., Hart, P. E., & Stork, D. G. (2001). Pattern classification (2nd

ed.). New York, USA: Wiley.

Dumont, J.-P., Roger, S., & Adda, J. (1974). Volatile compounds in whole

cheese and grated cheese: Parmesan. Lait, 54, 386–396.

EC (1996). Official Journal of the European Union, Regulation no. 1107/

1996 of 21 June 1996.

Engels, W. J. M., Dekker, R., de Jong, C., Neeter, R., & Visser, S. (1997).

A comparative study of volatile compounds in the water-soluble

fraction of various types of ripened cheese. International Dairy Journal,

7, 255–263.

Engels, W. J. M., & Visser, S. (1994). Isolation and comparative

characterization of components that contribute to the flavour of

different cheese types. Netherlands Milk and Dairy Journal, 48,

127–140.

Eriksson, L., Johansson, E., Kettaneh-Wold, N., & Wold, S. (1999).

Introduction to multi and megavariate data analysis using projection

methods (PCA & PLS). Umea, Sweden: Umetrics AB.

Forina, M., Lanteri, S., Armanino, C., Cerrato-Oliveiros, C., & Casolino,

C. (2003). V-Parvus 2003: An extendable package of programs for data

explorative analysis, classification and regression analysis. Department

of Chimica e Tecnologie Farmaceutiche e Alimentari, University of

Genova, Genova, Italy (free download at http://http.parvus.unige.it).

Fox, P. F., Singh, T. K., & McSweeney, P. L. H. (1995). Biogenesis of

flavour compounds in cheese. In E. L. Malin, & M. H. Tunick (Eds.),

Chemistry of structure/function relationships in cheese (pp. 59–98).

London, UK: Plenum Press.

Fox, P. F., & Wallace, J. M. (1997). Formation of flavour compounds.

Advances in Applied Microbiology, 45, 17–85.

Good, I. (1999). Resampling methods: A practical guide to data analysis.

Boston, MA, USA: Birkhauser.

Hansel, A., Jordan, A., Holzinger, R., Prazeller, P., Vogel, W., &

Lindinger, W. (1995). Proton transfer reaction mass spectrometry: On-

line trace gas analysis at the ppb level. International Journal of Mass

Spectrometry and Ion Processes, 149/150, 609–619.

Karl, T., Yeretzian, C., Jordan, A., & Lindinger, W. (2003). Dynamic

measurements of partition coefficients using proton-transfer-reaction

mass spectrometry (PTR-MS). International Journal of Mass Spectro-

metry, 223–224, 383–395.

Kemsley, E. K. (1998). Discriminant analysis and class modelling of

spectroscopic data. Chichester, UK: Wiley.

Kress-Rogers, E. (1996). Biosensors and electronic noses for practical

applications. In E. Kress-Rogers (Ed.), Biosensors and electronic noses:

medicine, food, and the environment. Frankfurt, Germany: CRC Press.

Kristoffersen, T. (1973). Biogenesis of cheese flavor. Journal of Agricultur-

al and Food Chemistry, 21, 573–575.

Lagg, A., Taucher, J., Hansel, A., & Lindinger, W. (1994). Application of

proton transfer reaction to gas analysis. International Journal of Mass

Spectrometry and Ion Processes, 134, 55–56.

Larrayoz, P., Addis, M., Gauch, R., & Bosset, J. O. (2001). Comparison of

dynamic headspace and simultaneous distillation extraction techniques

used for the analysis of the volatile components in three European PDO

ewes’ milk cheeses. International Dairy Journal, 11, 911–926.

Law, B. A. (1984). Flavour developments in cheeses. In F. L. Davies, & B.

A. Law (Eds.), Advances in the microbiology and biochemistry of cheese

and fermented milk (pp. 197–207). London, UK: Elsevier.

Lindinger, W., Hansel, A., & Jordan, A. (1998). On-line monitoring of

volatile organic compounds at ppt level by means of proton-transfer-

reaction mass spectrometry (PTR-MS): Medical application, food

control and environmental research. International Journal of Mass

Spectrometry and Ion Processes, 173, 191–241.

Lindinger, W., Hirber, J., & Paretzke, H. (1993). An ion/molecule-reaction

mass spectrometer used for on-line trace gas analysis. International

Journal of Mass Spectrometry and Ion Processes, 129, 79–88.

Liu, S.-Q., Holland, R., & Crow, V. L. (2004). Esters and their

biosynthesis in fermented dairy products: A review. International

Dairy Journal, 14, 923–945.

Maarse, H., & Visscher, C. A. (1989) (6th ed.). Volatile compounds in food.

qualitative and quantitative data. Zeist I, The Netherlands: Central

Institute for Nutrition and Food Research TNO-CIVO.

Manning, D. J. (1979). Chemical production of essential Cheddar cheese

flavour. Journal of Dairy Research, 46, 531–537.

Manning, D. J., Ridout, E. A., & Price, J. C. (1984). Non-sensory methods

for cheese flavour assessment. In F. L. Davies, & B. A. Law (Eds.),

Advances in the microbiology and biochemistry of cheese and fermented

milk (pp. 229–253). London, UK: Elsevier.

Marcos Lorenzo, I., Perez Pavon, J. L., Fernandez Laespada, M. E.,

Garcya Pinto, C., & Moreno Corsero, B. (2002). Detection of

adulterants in olive oil by headspace–mass spectrometry. Journal of

Chromatography A, 945, 221–230.

McSweeney, P. L. H. (1997). The flavour of milk and dairy products: III.

Cheese taste. International Journal of Dairy Technology, 50, 123–128.

McSweeney, P. L. H., & Sousa, M. J. (2000). Biochemical pathways for

the production of flavour compounds in cheese during ripening: A

review. Lait, 80, 293–324.

Meinhart, E., & Schreier, P. (1986). Study of flavour compounds from

Parmigiano-Reggiano cheese. Milchwissenschaft, 41, 689–691.

Moio, L., & Addeo, F. (1998). Grana Padano cheese aroma. Journal of

Dairy Research, 65, 317–333.

ARTICLE IN PRESSE. Aprea et al. / International Dairy Journal 17 (2007) 226–234234

Munson, M. S. B., & Field, F. H. (1966). Chemical ionization mass

spectrometry. I. General introduction. Journal of the American

Chemical Society, 88, 2621–2630.

Ott, A., Germond, J. E., & Chaintreau, A. (2000). Origin of acetaldehyde

during milk fermentation using 13C-labeled precursors. Journal of

Agricultural and Food Chemistry, 48, 1512–1517.

Pollien, P., Jordan, A., Lindinger, W., & Yeretzian, C. (2003). Liquid–air

partitioning of volatile compounds in coffee: Dynamic measurements

using proton-transfer-reaction mass spectrometry. International Jour-

nal of Mass Spectrometry, 228, 69–80.

Stone, M. (1974). Cross-validatory choice and assessment of statistical

predictions. Journal of the Royal Statistical Society. Series B

(Methodological), 36(2), 111–147.

Stratton, J. E., Hutkins, R. W., & Taylor, S. L. (1991). Biogenic amines in

cheese and other fermented foods: A review. Journal of Food

Protection, 54, 460–470.

Wold, S., Albano, C., Dum, W. J., Edlund, U., Esbensen, K., Geladi, P.,

et al. (1984). Multivariate data analysis in chemistry. In B. R.

Kowalski (Ed.), Chemometrics, mathematics and statistics in chemistry.

Dordrecht, Holland: Riedel Publishing.

Yeretzian, C., Jordan, A., & Lindinger, W. (2003). Analysing the

headspace of coffee by proton-transfer-reaction mass-spectro-

metry. International Journal of Mass Spectrometry, 223&224,

115–139.

Yvon, M., & Rijnen, L. (2001). Cheese flavour formation by amino acid

catabolism. International Dairy Journal, 11, 185–201.