Embed Size (px)
Citation preview
HAL Id: hal-00895684https://hal.archives-ouvertes.fr/hal-00895684
Submitted on 1 Jan 2008
HAL is a multi-disciplinary open accessarchive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come fromteaching and research institutions in France orabroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, estdestinée au dépôt et à la diffusion de documentsscientifiques de niveau recherche, publiés ou non,émanant des établissements d’enseignement et derecherche français ou étrangers, des laboratoirespublics ou privés.
Multivariate statistical methods for discrimination oflactobacilli based on their FTIR spectra
Dragiša Savić, Nataša Joković, Ljubiša Topisirović
To cite this version:Dragiša Savić, Nataša Joković, Ljubiša Topisirović. Multivariate statistical methods for discriminationof lactobacilli based on their FTIR spectra. Dairy Science & Technology, EDP sciences/Springer, 2008,88 (3), pp.273-290. <hal-00895684>
Dairy Sci. Technol. 88 (2008) 273–290 Available online at:c© INRA, EDP Sciences, 2008 www.dairy-journal.orgDOI: 10.1051/dst:2008003
Original article
Multivariate statistical methodsfor discrimination of lactobacilli based
on their FTIR spectra
Dragiša Savic1*, Nataša Jokovic1, Ljubiša Topisirovic2
1 Laboratory for Food and Biotechnology, Faculty of Technology, Bulevar oslobodjenja 124,16000 Leskovac, Serbia
2 Laboratory for Molecular Genetics of Industrial Microorganisms, Institute of Molecular Genetics andGenetic Engineering, Vojvode Stepe 444a, P.O. Box 23, 11010 Belgrade, Serbia
Received 2 October 2007 – Accepted 31 December 2007
Abstract – In this study, 41 strains belonging to 6 Lactobacillus species originating from culturecollections, and isolates from homemade cheese that had previously been genetically classified,were discriminated based on their FTIR spectra. The lactobacilli species L. brevis, L. diolivorans,L. plantarum and L. rhamnosus could be differentiated on the basis of their FTIR spectra by apply-ing multivariate statistical techniques (Hierarchical Cluster Analysis, Principal Component Anal-ysis and Multidimensional Scaling). The strains belonging to L. paracasei and L. casei could notbe separated from each other, and formed one group that was distanced from the other investigatedspecies. The combination of common multivariate statistical methods used for the analysis of FTIRspectra could provide a powerful method for discrimination and even the identification of lacto-bacilli on a species level. The strains and subspecies within the lactobacilli species could not bedifferentiated by applying multivariate exploratory statistical methods. The FTIR spectra of lacto-bacilli were robust against the variation in the consistency of the growth medium (broth or agarmedium of the same composition), applied spectroscopic techniques (KRS-5 or tablets) and therecorded biomass quantities of 2 and 3 mg. Furthermore, the distances between the two spectra ofthe same selected strain, which were sampled ca. 12 months apart, were insignificant.
lactobacilli / FTIR spectroscopy / multivariate statistical methods
摘摘摘要要要 –基基基于于于傅傅傅立立立叶叶叶变变变换换换红红红外外外光光光谱谱谱和和和多多多元元元统统统计计计分分分析析析方方方法法法鉴鉴鉴定定定乳乳乳杆杆杆菌菌菌。。。本文采用傅立叶变换红外光谱 (FTIR)法对源于手工制作干酪中的 6个种 41株乳酸菌进行了鉴定。运用多元统计分析方法 (聚类分析, 主成分分析和多维尺度分析) 对短乳杆菌 (L. brevis), L. diolivorans,植物乳杆菌 (L. plantarum)和鼠李糖乳杆菌 (L. rhamnosus)的 FTIR光谱数据进行了分析。但对其中副干酪乳杆菌 (L. paracasei)和干酪乳杆菌 (L. casei)不能有效地识别。 FTIR光谱技术结合多元统计分析方法可以将乳杆菌鉴定到种,但是应用这种探索性的多变量统计分析方法不能鉴定到株或亚种。通过研究发现乳杆菌的 FTIR光谱与生长条件 (在相同成分的肉汤或琼脂培养基)不同、使用的光谱制样技术 (KRS-5光滑面或压片式)以及用来记录的生物量 (2 mg和 3 mg)的不同而有显著的差异。此外, 保存期相差约 12个月的同菌株之间的光谱没有显著差异。
乳乳乳杆杆杆菌菌菌 /傅傅傅立立立叶叶叶变变变换换换红红红外外外光光光谱谱谱 (FTIR) /多多多元元元统统统计计计方方方法法法
* Corresponding author (通讯作者): [email protected]
Article published by EDP Sciences and available at http://www.dairy-journal.org or http://dx.doi.org/10.1051/dst:2008003
274 D. Savic et al.
Résumé – Méthodes statistiques multidimensionnelles pour la discrimination des lactoba-cilles basée sur leur spectre IRTF. Dans cette étude, 41 souches appartenant à six espèces deLactobacillus provenant de collections de cultures ou isolées de fromages faits maison, précé-demment classées génétiquement, sont discriminées selon leur spectre infrarouge à transforméede Fourier (IRTF). Les espèces Lactobacillus brevis, L. diolivorans, L. plantarum et L. rhamnosuspeuvent être différenciées à partir de leur spectre IRTF en appliquant des techniques statistiquesmultidimensionnelles (classification ascendante hiérarchique, analyse en composantes principaleset positionnement multidimensionnel). Les souches de L. paracasei et L. casei ne peuvent pas êtreséparées entre elles et forment un groupe distant des autres espèces étudiées. La combinaison deméthodes statistiques multidimensionnelles, couramment utilisées pour l’analyse de spectres IRTF,peut fournir une méthode puissante pour la discrimination et même l’identification de lactobacillesà l’échelle de l’espèce. Les souches et sous-espèces parmi les espèces de lactobacilles ne peuventpas être différenciées en appliquant des méthodes statistiques multidimensionnelles exploratoires.Les spectres IRTF des lactobacilles sont robustes quels que soient les changements de consistancedes milieux de culture (milieu liquide ou gélosé de la même composition), les méthodes spectro-scopiques appliquées (KRS-5 ou comprimé) ou la quantité de biomasse analysée (2 ou 3 mg). Deplus, la distance entre deux spectres d’une même souche sélectionnée, échantillonnée avec 12 moisd’écart, est insignifiante.
lactobacilles / spectroscopie IRTF / méthodes statistiques multidimensionnelles
1. INTRODUCTION
The great potential of Fourier Trans-form Infrared (FTIR) spectroscopy, beinga rapid, easy to use and cost-effective tech-nique, has been recognized in discrimi-nation, classification and identification ofbacteria, yeast and other microorganisms(for a review, see [14]). The cells are mea-sured in a non-destructive manner and theinfrared spectra of the microorganism stud-ied are considered as its chemical finger-print. The FTIR spectrum is given by thevibrational properties of all the cell con-stituents, including fatty acids, proteins,polysaccharides and nucleic acids [2]. Adatabase with the reference microorgan-isms’ spectra has to be established, and theunknown microorganisms can be identifiedby comparing their FTIR spectra with thespectra in the database.
The most commonly used FTIR tech-nique applied so far for recording spectrain studies of microorganism identificationis a transmission by placing the sample ona ZnSe optical plate [1,2,4,16–18,22]. Be-sides this, spectroscopic studies using thesamples in the form of tablets have alsobeen used for discrimination of species
within the genera Listeria [12] and Bru-cella [5]. The samples were prepared bycollecting the biomass from agar plates[6, 8, 16–18, 22] or a centrifuged broth cul-ture [1, 5, 10, 12, 13, 20].
Since the amount of the data obtainedis large, the differences between the spec-tra of the microorganisms are described byusing statistical methods. Many statisticaltechniques have been extensively appliedin the analysis of the infrared spectra in mi-croorganism discrimination: Partial LeastSquares Regression [16–18], Soft Indepen-dent Modeling of Class Analogy [10, 17],K-Nearest Neighbor [10, 17] and Artifi-cial Neural Networks [6]. However, mul-tivariate exploratory techniques, such asHierarchical Cluster Analysis [2, 4, 8, 17],Principal Component Analysis [5, 10, 17],Discriminant Analysis [1, 22] and Canon-ical Variate Analysis [12] have been themost common and most often applied sta-tistical methods. So far, there have beenno studies on the use of Multidimen-sional Scaling (MDS) analysis for differ-entiation of microorganisms based on theirFTIR spectra. In general, the goal of MDSanalysis is to detect meaningful underly-ing dimensions that allow the researcher
Discrimination of lactobacilli by FTIR 275
to explain the observed similarities ordissimilarities (distances) between the ob-jects investigated. Recently, MDS wasused to obtain a bidimensional represen-tation of the position and the diversity ofbacterial communities based on their phe-notypic properties [19].
Since lactic acid bacteria (LAB) arewidely used in the food industry, especiallyas starters in fermentation processes, therapid and reliable identification and char-acterization of these microorganisms areimportant. There are few reports on theanalysis of the relationships and the identi-fication of LAB using FTIR spectroscopy.Oust et al. [17] analyzed 56 strainsfrom four Lactobacillus species (L. sakei,L. plantarum, L. curvatus and L. paraca-sei). By applying FTIR analysis, the au-thors succeeded in recognizing the strainsthat were incorrectly identified by conven-tional methods prior to the FTIR analy-sis. Likewise, FTIR spectroscopy correctlyidentified 91% of the facultatively hetero-fermentative lactobacilli previously iden-tified with SDS-PAGE, with high scoresfor L. plantarum (100% identification) andL. rhamnosus (91%), and a relatively lowscore for L. casei (87%) [22]. The re-sults of Lefier et al. [13] demonstratedthat FTIR was a rapid and robust methodfor a qualitative analysis of the density oflactococcal flora during soft cheese matu-ration. The identification of 100% at thegenus level and 69% at the species levelfor LAB isolates from the cheese industrywas established by using FTIR [1]. More-over, 53 strains from 14 Lactobacillusspecies found in breweries were analyzedby FTIR spectroscopy and identification atthe species level was correct for 94% of thestrains, and at the strain level for 91% ofthe strains [2].
The main purpose of this study wasto discriminate and to find the relationsamong Lactobacillus isolates from home-made cheeses that had previously been ge-netically classified, and species originating
from culture collections, on the basis oftheir FTIR spectra. The quality of discrim-ination was evaluated by comparing theidentification results from commonly usedmultivariate exploratory statistical meth-ods and, for the first time, a Multidi-mensional Scaling Analysis. The influenceof the growth conditions (the agar plateand broth of the same composition), thespectroscopic sampling technique applied(a transmission through tablets or opticalplate), recorded biomass quantity and stor-age by freezing on lactobacilli discrimina-tion was also examined.
2. MATERIALS AND METHODS
2.1. Strains and growth conditions
Forty-one Lactobacillus strains of dif-ferent origin (Tab. I) were analyzed withFTIR spectroscopy. The strains under in-vestigation were reference strains fromculture collections and isolates from home-made cheeses that had previously beengenetically classified [9]. Four strainswere L. brevis (the same strain type ob-tained from two collections, LMG 6906and NRRL B-4527, was the subject ofthe study), 5 strains were L. diolivorans,2 strains were L. casei, 7 were L. paraca-sei (2 subspecies), 20 strains were L. plan-tarum (2 of which belonged to 2 sub-species) and 3 were L. rhamnosus. Thenomenclatures of the strains used are citedfrom the culture collections, and the taxo-nomic status of the species and the strainsat issue is not discussed.
The strains were cultivated on MRSagar and MRS broth (Torlak, Belgrade) at30 ◦C for 48 ± 1 h. A platinum loop wasused to inoculate the agar plate (streakingout) and the broth. Three sampling replica-tions were done for each culture.
276 D. Savic et al.
Table I. Lactobacilli strains used in this study.
Species Strain Isolation environmentL. brevis NRRL1 B-4527T Feces, human
(LMG 6906T)LMG 7761 Green fermenting olivesLMG 11401 BeerLMG 18022 Zabady (yogurt) Burundi
L. diolivorans BGGJ 7-18 CheeseBGGJ 7-24BGGJ 7-27BGGJ 8-25BGGJ 8-13
L. casei NRRL B-1922T CheeseNRLL B-441 Cheese, Emmental
L. paracasei ssp. paracasei NRRL B-4560T Milking machineNRRL B-4564T Saliva, childLMG 11459 Dental cariesBGBUK2-16/K2 CheeseCI 5-25* 20-day ripened cow’s cheeseCI 2-21* Ewe’s milk
L. paracasei ssp. tolerans LMG 9191T Pasteurized milkL. plantarum NRRL B-4496T Pickled cabbage
LMG 18024 Buffalo milk, EgyptLMG 18023 Mish cheese, EgyptLMG 9212 Italian human, salivaBG A112 CheeseCI 4-3* 10-day ripened cow’s cheeseCI 6-8* 30-day ripened cow’s cheeseCI 6-9*CI 6-11*CI 6-16*CI 6-28*CI 7-12* 60-day ripened cow’s cheeseCI 7-16*CI 7-20*CI 7-21*CI 7-22*CI 7-31*CI 7-32*
L. plantarum ssp. argentoratensis LMG 9205T Fermented corn productL. plantarum ssp. rudensis LMG 11475 Cheese showing rusty spotL. rhamnosus NRRL B-442T –
BGZL-19 CheeseBGHV-747 Vaginal isolate
1 NRRL: Agricultural Research Service Culture Collection, National Center for AgriculturalUtilization Research, Peoria, Illinois, USA.LMG: Bacteria Collection, Laboratorium voor Microbiologie, Universiteit Gent, Belgium.BG: Bacteria Collection, Institute of Molecular Genetics and Genetic Enginering, Belgrade, Serbia.T Marked as Type strain in the Culture Collection.* CI: Isolates from cheese determined by Rep-PCR [9].
Discrimination of lactobacilli by FTIR 277
2.2. Sample preparation for FTIRspectroscopy
After the cultivation on agar plates, ametal spatula was used to remove thebacterial biomass. The cells grown inbroth were separated by centrifugation(5000 min−1 for 15 min), washed twicewith distilled water and dried at 60 ◦C for60 min under a moderate vacuum.
Optical plate measurement
Preliminary trials were performed to de-termine the density of the cell suspensions,necessary to produce spectra with a goodsignal to noise ratio without band satura-tion. The required densities were obtainedby transferring 5 mg of collected biomassfrom an agar plate to a microtube (1.5 cm3)and suspending it in 0.250 cm3 of dis-tilled water. 0.1 cm3 (containing ca. 2 mgof biomass) of the aliquot was placed ona KRS-5 optical plate (25 × 4 mm, DSCpol. Crystal Labs, NY), and dried under amoderate vacuum in a desiccator to formfilm suitable for FTIR measurement. Sinceprevious studies showed that drying of be-tween 15 and 45 min had no influence on aFTIR spectrum (data not shown), the sam-ples were desiccated for ca. 20 min.
Tablet preparation
The biomass collected from dried sed-iments or the agar plate (ca. 2 or 3 mg)was mixed and ground with 150 mg KBr(Merck, Germany). This powder was thencompressed into a thin KBr disc (∅ 13 mm)under a pressure of 200 MPa for about5 min under a moderate vacuum.
2.3. FTIR spectroscopy
All the spectra were recorded in theregion of between 4000 and 400 cm−1
using a BOMEM MB-100 FTIR Spectrom-eter (Hartmann & Braun, Quebec, Canada)equipped with a DTGS detector model
D31B. For data acquisition, the softwareWin-Bomem Easy 3.1 was used. For eachsample recording (through tablets or opti-cal plate), 10 scans were averaged with aspectral resolution of 4 cm−1.
2.4. Preprocessing of the spectra
The IR spectra obtained by recordingthe samples on the KRS-5 optical plateor in the form of tablets were in accor-dance with the quality control test basedon the criteria of a signal-noise ratio andsignal water relations [4]. In order to re-move variations due to baseline shifts, thefirst derivative of every spectrum was cal-culated using a Savitzky-Golay algorithmwith nine smoothing points [8, 17].
2.5. Data analysis
All statistical analyses were performedusing the STATISTICA 7.0 programfor Windows (StatSoft Inc., USA). Theprocessed spectra of the strains weretaken as variables (columns of the inputmatrix) and wavenumbers as statisticalcases (rows of the matrix). Multivariateexploratory techniques, HierarchicalCluster Analysis, Principal ComponentAnalysis and Multidimensional ScalingAnalysis, were performed to compare theresults of lactobacilli discrimination. As itwas found to be best suited for lactobacilliidentification [1, 2, 17, 18], the regionbetween 1400 and 700 cm−1 was used fordiscrimination of the species.
Hierarchical Cluster Analysis (HCA)was used to measure the similaritiesand to classify the lactobacilli spectrainto groups. The distances were calcu-lated using Pearson’s product momentcorrelation coefficient, and clusteringwas carried out using Ward’s algorithm.The Statistica software calculates and
278 D. Savic et al.
transforms Pearson’s correlations intodissimilarities (1 – Pearson r), meaningthat a value of 1.00 represents a lack ofcorrelation, whereas a value of 0.00 repre-sents a perfect correlation.
Principal Component Analysis (PCA)was performed by using the algorithmknown as NIPALS (Non-linear IterativePartial Least Squares). The NIPALSalgorithm is an iterative procedure usedfor constructing PCA models for the rep-resentation of multivariate variables usinga number of principal components that areless in dimension than a number of theoriginal variables. It is the most commonlyused method for calculating the principalcomponents, and it provides numericallymore accurate results as compared withother methods [21].
Multidimensional Scaling (MDS) istypically used as an exploratory method.To discriminate the spectra of the lacto-bacilli used in this study, the matrix of thePearson’s product moment correlation co-efficients was computed and submitted toMDS.
2.6. Experimental setup
Experiments were performed in twoparts. The first part of the experiment in-volved the assessment of the importanceof the recording technique. The spectraof eight strains, representing the speciesstudied (L. plantarum NRRL B-4496 andCI 7-12, L. plantarum ssp. argentoraten-sis LMG 9205, L. brevis NRRL B-4527,L. diolivorans BGGJ 8-25, L. casei NRRLB-441, L. paracasei ssp. paracasei LMG11459 and L. rhamnosus NRRL B-442)were obtained by using two transmissiontechniques, placing the samples both astablets and on the KRS-5 optical plate.To provide approximately the same quan-tity of biomass for FTIR spectra recording,
ca. 2 mg of the biomass collected from theagar plate was used both for tablet prepa-ration and for placing on the KRS-5 plate(0.1 cm3 suspension of ca. 5 mg biomassfrom the agar plate in 0.250 cm3 of distilledwater). In order to determine the effects ofthe biomass concentration on FTIR spec-tra, the tablets were also prepared with ca.3 mg of the collected biomass. Addition-ally, spectra were recorded by using tabletsprepared with ca. 3 mg of the biomass col-lected from the cells grown in MRS broth.
In the second part, the spectra of all therest of the reference strains and the iso-lates from cheese (Tab. I) were recordedusing the tablets that were prepared withca. 3 mg of the biomass collected from theagar plate. Finally, the spectra of all thestrains were recorded ca. 12 months apartafter storing bacteria by freezing (–18 ◦C)in a mixture of glycerin (15%) and MRSbroth. After storing, the spectra of the bac-teria were recorded by applying the tablettechnique (with ca. 3 mg biomass), after re-viving them by subculturing twice on theMRS agar (48 h, 30 ◦C).
3. RESULTS
During the FTIR measurement, a typicalspectrum of lactobacilli with various spec-tral regions was obtained (data not shown)as described earlier [17]. In data analy-sis, only the parts of the spectra between1400 and 700 cm−1, which cover the mixedregion, the polysaccharide region and the“true” [17] fingerprint region, were appliedfor statistical analysis, since that regionwas found to be best suited for differen-tiation and the identification of the Lacto-bacillus species [1, 2, 17, 18]. The absorp-tion bands at ∼1080 cm−1 and ∼1242 cm−1
assigned to symmetric and asymmetricstretching vibration of phosphate func-tional groups of nucleic acids, respec-tively [10], dominated the FTIR spectra re-gion of between 1400 and 700 cm−1 of
Discrimination of lactobacilli by FTIR 279
Figure 1. The 1400–700 cm−1 region of FTIR spectra (KBr tablets, 3 mg of biomass) of selectedLactobacillus strains cultivated on MRS agar (30 ◦C for 48 ± 1 h).
all the selected lactobacilli (Fig. 1). Thesignificant differences in the raw spectra ofthe selected lactobacilli could be observedin the range of 1200–950 cm−1.
3.1. Effects of growth conditions,spectroscopic techniqueand biomass concentration
In preliminary studies (experimentpart I) eight strains, representing the
species studied (L. plantarum NRRLB-4496 and CI 7-12, L. plantarum ssp.argentoratensis LMG 9205, L. brevisNRRL B-4527, L. diolivorans BGGJ8–25, L. casei NRRL B-441, L. paracaseissp. paracasei LMG 11459 and L. rham-nosus NRRL B-442) were grown on MRSagar plates. The biomasses were collectedfrom the agar plates and the spectra wererecorded by applying two spectroscopictechniques (KRS-5 plate and tablets). Tocompare data, approximately the same
280 D. Savic et al.
quantity of the biomass was put both onthe KRS-5 plate and in tablets (ca. 2 mg).Since it was very difficult to measure smallquantities, such as 2 mg, precisely, thetablets were prepared with ca. 2 or 3 mgof the biomass collected from the platein order to determine the influence of thebiomass concentration in that range on thespectra differentiation. In order to examinethe possibility of including the lactobacillispectra that grew poorly on the MRS agar(could not yield 2–3 mg of biomass perplate) in discrimination studies, the effectsof the spectra of the cells grown in MRSbroth of the selected strains were alsoincluded in this study.
For all the eight strains, it was shownthat the effects on the spectra caused bygrowth in either broth or agar plate, appliedspectroscopic technique and the biomassquantity, were so small that the separationof the strains was nearly unaffected. Thiswas illustrated in the distances (expressedas 1 – Pearson’s correlation coefficients)between the strains and within the strain(Tab. II). However, the average spectra dis-tances within the same strain caused byvariation in experimental conditions var-ied in the range of 0.01 to 0.04. The av-erage distances between the spectra of thestrains belonging to different species weresignificantly larger than spectral distancesbetween the FTIR spectra of the samespecies. An exception was noticed in thecase of L. casei NRRL B-441 and L. para-casei ssp. paracasei LMG 11459, wherethe interspecies distances were equal to in-traspecies distances (0.03). Also, the spec-tra of the reference strains L. plantarumNRRL B-4496 and L. plantarum ssp. ar-gentoratensis LMG 9205 showed a highlevel of interspecies similarity (spectra dis-tance 0.03), but the spectrum of the iso-late from cheese, L. plantarum CI7-12, wassignificantly distanced from them (spectradistances 0.08 and 0.10) (Tab. II).
For further studies, a technique ofpreparing tablets with ca. 3 mg of the
biomass was selected despite the opin-ion [6] that preparing a sample as tabletswas a slow and rather tedious method andreplacement with the application of liq-uid samples was recommended. Preparingtablets was selected because it was fasterand it allowed continuous laboratory work.
3.2. Discrimination of species
In order to determine the distances be-tween the FTIR spectra of the species com-monly found in cheeses, and to test a statis-tical approach, HCA was first performed todifferentiate the average spectra of 26 ref-erence strains from the culture collections(LMG, NRRL and BG). The dendrogramfrom the HCA of the FTIR spectra of the7 reference strains (experiment part I) andthe spectra of all the rest of the referencestrains (recorded before and after storingas tablets with ca. 3 mg biomass fromagar plates) was constructed by applyingPearson’s product moment correlation co-efficient for a distance measurement andWard’s algorithm as aggregation algorithm(Fig. 2A). Pearson’s correlation coefficientwas chosen since it is a common parameterused to quantify the strength of the linearassociation between variables [2, 8, 20], aswell as Ward’s algorithm, which is widelyused and well suited for the FTIR spectraof microorganisms [14].
The reference strains L. brevis (inter-species linkage distance ca. 0.2), L. dio-livorans (interspecies linkage distance ca.0.1), L. rhamnosus (interspecies linkagedistance ca. 0.1) and L. plantarum (inter-species linkage distance ca. 0.3) clusteredaccording to the species description re-gardless of their origin. Even L. plantarumssp. argentoratensis (LMG 9205) andL. plantarum ssp. rudensis (LMG 11475)were not separated from other L. plan-tarum species (Fig. 2A). The HCA tech-nique applied could not separate L. para-casei ssp. paracasei, L. paracasei ssp.
Discrimination of lactobacilli by FTIR 281
Tabl
eII
.Dis
tanc
em
easu
rem
ent
aver
age
(1–
Pea
rson
’sr1
)an
dst
anda
rdde
viat
ion
(in
pare
nthe
sis)
betw
een
FT
IRsp
ectr
a2of
the
stra
ins
repr
esen
ting
the
lact
obac
illi
stud
ied.
L.p
lant
arum
L.b
revi
sL
.dio
livor
ans
L.c
asei
L.p
arac
asei
L.r
ham
nosu
sSt
rain
ssp.
arge
n.N
RR
LC
IL
MG
NR
RL
BG
GJ
NR
RL
LM
GN
RR
LB
-449
67-
1292
05B
-452
78-
25B
-441
1145
9B
-442
L.p
lant
arum
0.01
NR
RL
B-4
496
(0.0
06)
CI
7-12
0.08
0.04
(0.0
3)(0
.01)
ssp.
arge
nora
tens
is0.
030.
100.
02L
MG
9205
(0.0
1)(0
.03)
(0.0
1)L
.bre
vis
0.10
0.10
0.11
0.02
NR
RL
B-4
527
(0.0
2)(0
.02)
(0.0
3)(0
.006
)L
.dio
livor
ans
0.11
0.14
0.11
0.08
0.02
BG
GJ
8-25
(0.0
2)(0
.01)
(0.0
2)(0
.02)
(0.0
1)L
.cas
ei0.
080.
130.
050.
140.
080.
03N
RR
LB
-441
(0.0
1)(0
.02)
(0.0
1)(0
.02)
(0.0
2)(0
.01)
L.p
arac
asei
ssp.
para
case
i0.
090.
110.
060.
100.
050.
030.
01L
MG
1145
9(0
.02)
(0.0
1)(0
.01)
(0.0
2)(0
.01)
(0.0
1)(0
.007
)L
.rha
mno
sus
0.12
0.17
0.09
0.19
0.12
0.08
0.08
0.02
NR
RL
B-4
42(0
.02)
(0.0
2)(0
.02)
(0.0
2)(0
.02)
(0.0
1)(0
.01)
(0.0
1)
1T
heS
tati
stic
aso
ftw
are
calc
ulat
esan
dtr
ansf
orm
sco
rrel
atio
nsin
todi
ssim
ilar
itie
s(1
–P
ears
onr)
[21]
.2
For
each
stra
in,f
our
aver
aged
spec
tra
ofth
ree
inde
pend
ent
sam
ples
wer
eob
tain
ed:
1on
aK
RS
-5op
tica
lpl
ate
(gro
wn
onM
RS
agar
)an
d3
inth
efo
rmof
tabl
ets
(1w
ith
ca.2
mg
and
1w
ith
ca.3
mg
ofce
lls
grow
non
MR
Sag
ar,a
nd1
wit
hca
.3m
gbi
omas
sco
llec
ted
from
cell
sgr
own
inM
RS
brot
h).T
heva
lues
inth
eta
ble
wer
eca
lcul
ated
asav
erag
edi
stan
ces
betw
een
each
ofth
est
rain
s’4
spec
tra.
The
bold
valu
esre
pres
enta
vera
gedi
stan
ces
betw
een
the
spec
tra
ofth
esa
me
stra
inob
tain
edby
4diff
eren
tmet
hods
.Zer
ova
lues
betw
een
the
sam
esp
ectr
aw
ere
excl
uded
from
calc
ulat
ions
.
282 D. Savic et al.
Figure 2. Dendrograms from HCA of the average spectra of lactobacilli investigated sampledca. 12 months apart. The samples were recorded as tablets with ca. 3 mg of biomass collectedfrom an agar plate before (suffix T) and after storing the bacteria by freezing (suffix S). The dis-tances were calculated using Pearson’s product moment correlation coefficient and grouping wasperformed by Ward’s algorithm. (A) Dendrogram of strains from the culture collections includingspectra of 7 reference strains recorded under different experimental conditions: on a KRS opticalplate (suffix P), as tablets with ca. 2 mg of biomass collected from an agar plate (suffix t) and astablets with biomass collected from broth by centrifugation (suffix B).* Strain L. brevis LMG 6906 was not recorded after storing since it was the same strain asNRRL B-4527.
Discrimination of lactobacilli by FTIR 283
Figure 2. Continued. (B) Dendrogram of spectra of all lactobacilli investigated.
284 D. Savic et al.
tolerans and L. casei, since those strainsformed one cluster corresponding to thelinkage distance of 0.1 (Fig. 2A). So, withthe exception of L. paracasei and L. ca-sei strains, the reference strains were sat-isfactorily discriminated on the basis oftheir FTIR spectra regardless of the spec-troscopic technique applied (KRS-5 ortablets), the growth medium consistency(agar or broth) and the quantity of therecorded biomass (2 or 3 mg per sam-ple). Furthermore, the strains sampled ca.12 months apart (before and after stor-ing the bacteria by freezing) were assignedto their species. The only exception wasL. brevis LMG 11401 which, after beingstored, was shown to be closer to the L. di-olivorans strains (Fig. 2A).
In order to find out if multivariate sta-tistical methods could be used to discrimi-nate and to identify lactobacilli of differentorigin at the species level based on FTIRspectroscopy, the isolates from homemadecheese that had previously been geneticallyclassified were analyzed together with thereference strains. The average spectra ofall the investigated lactobacilli obtainedand recorded under the same conditions(recorded before and after storing as tabletswith ca. 3 mg biomass from agar plates)were analyzed with HCA, PCA and MDS.
The isolates from homemade cheese fellinto already defined clusters by introducingthem into HCA (Fig. 2B). When the FTIRspectra of 13 identified L. plantarum iso-lates from cheese were analyzed togetherwith 7 reference strains, they increasedthe interspecies distances from ca. 0.3 toca. 0.9. This is probably due to differencesamong the isolates from cheese and the ref-erence L. plantarum strains already noticedin part I of the experiment, where spectraof the strain CI7-12 were distanced fromthe reference strains L. plantarum NRRLB-4496 and L. plantarum ssp. argeno-ratensis LMG 9205 (Tab. II). However, allL. plantarum strains were clustered into
one group that was differentiated from theother investigated lactobacilli.
On the other hand, the introduction of2 isolates, which were previously geneti-cally classified as L. paracasei ssp. para-casei, did not significantly influence theinterspecies distances in the L. paraca-sei and L. casei group. All the L. plan-tarum isolates seemed to be most relatedto L. plantarum ssp. rudensis LMG 11475,which they clustered with on a linkagedistance of 0.4 (Fig. 2B). Furthermore,a dendrogram showed that the spectra ofall the strains investigated, which wererecorded ca. 12 months apart (before andafter storing by freezing), clustered into thegroup corresponding to the species. Whenthe number of strains was increased inHCA, the previous misgrouping of L. bre-vis LMG 11401 (Fig. 2A) was correctedand the strain sampled after storing was as-signed to the L. brevis cluster (Fig. 2B). Inthis case, introducing more strains of thesame species in a statistical analysis helpedto achieve more appropriate differentiationamong the lactobacilli.
The clusters formed by HCA analysis(Fig. 2) could be clearly observed on theplots from the PCA and the MDS (Figs. 3and 4). Each cluster of the strains wasrepresented by an ellipse for its samples.This type of ellipse was determined insuch a way that the length of its horizon-tal and vertical projection onto the x- andy-axis (respectively) was equal to the mean± (range 0.95), where the mean and therange refer to the X or Y variable [21].
PCA discriminated all lactobacilli,yielding 4 Principal Components (PC)that explained 96.9% of the total variancein the data (data not shown). The firstPrincipal Component (PC1 accounted for89.8% of the total variance) correlatedwell (correlation coefficients between 0.88and 0.98, data not shown), with spectra ofall the strains investigated. So, a physicalmeaning of PC1 could be attributed to thespectra of all lactobacilli. PC2 (accounting
Discrimination of lactobacilli by FTIR 285
Figure 3. Principal component loadings from the PCA analysis of the reference strains’ FTIR spec-tra (open symbols) and isolates that had previously been genetically classified (closed symbols).The explained variance by PC2 and PC3 were 4.2% and 1.6%, respectively. Numbers representclusters from Figure 2.L. brevis ( ), L. plantarum (�), L. casei (∇), L. paracasei (Δ), L. diolivorans (♦), L. rhamnosus (◦).
for 4.2% of the total variance) separatedL. plantarum from all other investigatedlactobacilli, while PC3 (1.6% of thetotal variance) separated L. brevis andL. diolivorans from L. rhamnosus and thegroup of L. paracasei and L. casei strains(Fig. 3).
To evaluate if MDS was able to pro-vide a representation of the similarity re-lationships among the strains, which wascoherent with that obtained by HCA andPCA, the matrix of the data obtained fromthe spectra was used for the calculation ofthe Pearson’s product moment correlationcoefficient for the distance measurementamong the strains and then for MDS in twodimensions. The plot obtained by MDS(Fig. 4) was related to the dendrogramfrom HCA (Fig. 2). As shown in Figure 4,distances between strains belonging to thesame cluster were not necessarily lowerthan distances between strains belongingto different clusters, despite the fact thatthe same distance metrics (Pearson) were
used for both statistical procedures, mean-ing that HCA and MDS arranged interclus-ter distances in a different manner.
The relationship between the positionsof the ellipses obtained in plots from PCAand MDS (Figs. 3 and 4) and the posi-tion of the clusters identified with HCA(Fig. 2) was evident. The distances and thedegree of overlapping among the ellipsesindicated the relations among the strainsinvestigated and the efficiency of a statis-tical technique at separating the clusters.It was obvious that lactobacilli belongingto L. casei and L. paracasei species couldnot be separated by applying multivariatestatistical techniques to their FTIR spectra.As shown in Figure 4, MDS separated allthe clusters without overlapping of the el-lipses. The PCA could not clearly separatethe L. rhamnosus, L. paracasei and L. ca-sei clusters since there was a significantoverlapping between ellipses representingthese two clusters (Fig. 3).
286 D. Savic et al.
Figure 4. Bidimensional representation of the reference strains’ FTIR spectra (open symbols)and isolates that had previously been genetically classified (closed symbols) obtained by MDS ofthe matrix of Pearson’s product moment correlation coefficients. Numbers represent clusters fromFigure 2.L. brevis ( ), L. plantarum (�), L. casei (∇), L. paracasei (Δ), L. diolivorans (♦), L. rhamnosus (◦).
The L. plantarum cluster ellipse, whichrepresented 20 strains included in thestudy, was the largest (Figs. 3 and 4). How-ever, the size of an ellipse was not neces-sarily proportional to the number of differ-ent strains indicated. In fact, the ellipse of3 strains of L. rhamnosus species was dis-proportionately larger compared with theellipse of 5 L. diolivorans strains (Figs. 3and 4). This might reflect not only theinterspecies diversity (all L. diolivoransstrains were isolated from a single sam-ple), but also the statistical approach usedto evaluate the diversity.
From the plots obtained from PCA andMDS (Figs. 3 and 4), it can be seenthat one point representing the L. plan-tarum strain is significantly distanced fromthe L. plantarum ellipse and shown to becloser to the cluster representing L. para-casei and L. casei strains. This point rep-resented L. plantarum ssp. argentoratensisLMG 9205 recorded as tablets with 3 mgbiomass collected from the agar plate. Thesame strain sampled after ca. 12 months
of storage was assigned into the correctspecies (L. plantarum) ellipse. The sepa-ration of this sample was not observed byapplying HCA (Fig. 2).
4. DISCUSSION
According to a well-defined fact ofbuilding libraries with the reference spec-tra and standardization of the cultivationconditions, in order to ensure reproducibleresults as a prerequisite for using FTIRspectroscopy as a means for identificationof lactobacilli, more strains and specieshave to be included in the library. In thispaper, 41 lactobacilli strains belonging tosix species commonly found in cheesewere examined in order to help estab-lish standard conditions for cultivating andstandardizing a spectroscopic techniquefor recording FTIR spectra and findinga suitable statistical method for discrim-ination of lactobacilli. The results of ap-plying statistical multivariate techniques
Discrimination of lactobacilli by FTIR 287
showed that FTIR spectra of lactobacilliwere robust against a variation in the con-sistency of the growth medium (broth oragar medium of the same composition), theapplied spectroscopic techniques (KRS-5or tablets) and the biomass concentrationsof ca. 2 and ca. 3 mg per sample. Fur-thermore, all spectra of the investigatedstrains, which were sampled ca. 12 monthsapart (before and after storing by freezing),clustered into groups corresponding to thespecies.
The applied multivariate statisticalmethods appeared to be very effectivefor discrimination of the four speciesL. brevis, L. diolivorans, L. plantarum andL. rhamnosus based on their FTIR spectra.In fact, the clusters obtained by HCA, andthe position and the shape of the ellipsesobtained by PCA and MDS were related tothe species, and plotting the position of theclusters of the species on the same graphhelped identify this relationship.
In previous studies of lactobacilli iden-tification by applying HCA to their FTIRspectra, the data sets consisted of fewstrains from each of many species. Withthe exception of L. diolivorans, lacto-bacilli of all other species used in the cur-rent study were among 14 studied speciesisolated from beer in the research ofCurk et al. [2]. Despite some differences,good discrimination could be observed be-tween the lactobacilli species, as well as ahigh degree of relatedness between L. ca-sei and L. paracasei strains. On the otherhand, Oust et al. [17] showed that HCAwas not suited for the analysis of FTIRspectra of microorganisms with respect tospecies differentiation since it could notseparate lactobacilli on the species levelwhen the number of strains was increased.
PCA modeling is used to replace a com-plex multi-dimensional data set, such asinfrared spectra, by a simplified versioninvolving fewer dimensions which facil-itates an easier analysis of the variancein the data set [10]. In addition, in or-
der to reduce the dimensions of the orig-inal space of variables, PCA can alsobe used as a classification technique, sothat the relations among the variables andcases can be highlighted [21]. To discrimi-nate closely related lactobacilli, Oust et al.[17] used Partial Least Squared Regres-sion (PLSR), a statistical method that ex-tracts PLS components from a data matrix.It was shown that PC1 separated L. plan-tarum from three other studied species(L. sakei, L. curvatus and L. paracasei),PC2 separated L. paracasei, while PC3separated L. curvatus from the rest ofthe species. The results were not directlycomparable with the results in this paperbecause of the application of a differentstatistical technique for extracting compo-nents and a different model of data prepro-cessing. In the work of Oust et al. [17], thefirst derivation, vector normalization andweighting spectral data were used prior tostatistical analysis. In the present study,the normalization of the spectra was notdone, although it is a very often recom-mended and used method for vector or in-tegral spectra normalization [10,14,17,18].The method of normalization was not ap-plied since the first derivative was the mea-surement of a curve slope, so normaliza-tion, i.e. moving spectra up or down, didnot influence the first derivative. Further-more, attempts at the additional weight-ing of spectral data by dividing them bythe standard deviation of the data for eachwavenumber, as recommended earlier [17],did not give satisfactory discrimination(data not shown). So, on the basis of thisstudy and the results of Oust et al. [17], itcan be concluded that statistical techniqueswhich extract principal components fromFTIR spectra make it possible for L. plan-tarum to be clearly separated from otherlactobacilli.
According to the distances betweenthe ellipses that represent clusters of theinvestigated strains, MDS showed bettergrouping (Fig. 4) and discrimination of the
288 D. Savic et al.
investigated species than PCA (Fig. 3). Interms of resultant differences, PCA tendsto extract more components (dimensions)than MDS; as a result, MDS often yieldsmore readily interpretable solutions [21].Therefore, the lactobacilli were better dis-criminated in the first two dimensions(Fig. 4) by using MDS analysis, but not byimplementation of 3 or 4 components fromPCA. MDS is not so much an exact proce-dure as rather a way to “rearrange” the ob-jects in an efficient manner, so as to arriveat a configuration that best approximatesthe observed distances. The program actu-ally moves the objects around in the spacedefined by the requested number of dimen-sions, and checks how well the distancesbetween the objects can be reproduced bythe new configuration. In more technicalterms, the program uses a function mini-mization algorithm that evaluates differentconfigurations with the goal of maximizingthe goodness-of-fit (or minimizing “lack offit”) [21].
By introducing the previously genet-ically classified isolates into the study,FTIR spectroscopic analysis came to thesame result as a genomic identificationanalysis. In this way, genetic determina-tions of the isolates from cheese were con-firmed by comparing their FTIR spectrawith the spectra of the reference strains.This indicates that the reproducibility ofthe methods is good, which is importantif FTIR spectroscopy is to be used forthe identification of Lactobacillus at thespecies level. However, the average spec-tra distances of the same strain caused byvariation in experimental conditions werenot always lower compared with distancesbetween the strains belonging to the samespecies, regardless of the statistical methodapplied. So, the strains and subspecieswithin the species cannot be differentiatedby applying multivariate exploratory statis-tical methods to their FTIR spectra.
FTIR spectroscopy is not a reliable pa-rameter for establishing taxonomic rela-
tionships between microorganisms [15],and the results of this study confirmdifficulties in discrimination of phenotyp-ically related species, such as L. rhamno-sus, L. casei and L. paracasei [7]. By ap-plying multivariate statistical methods toFTIR spectra, highly related strains be-longing to L. paracasei ssp. paracasei,L. paracasei ssp. tolerans and L. caseicannot be separated from each other andform one group that is distanced from theother investigated species. The systemat-ics of the L. casei and L. paracasei grouphave been problematic [3, 11] and adjust-ments have been introduced based on newtaxonomic methods that have gradually be-come available. In this paper, the taxo-nomic status of the species and the strainsused was accepted as suggested in the re-spective culture collections. In addition, itseems that the L. rhamnosus strain canbe separated from the related group ofL. casei and L. paracasei by applying com-mon multivariate statistical methods, espe-cially HCA and MDS, but it is obvious thatmore strains of these species have to be in-cluded in making libraries of FTIR spec-tra. Generally, the identification results forall lactobacilli species will probably be im-proved if more strains of different originare included in the study.
Using HCA, PCA and MDS, the dis-crimination of all species was satisfactory,but there was at least one exception in eachstatistical approach. HCA misgrouped aL. brevis strain sampled after storage(Fig. 2A), while PCA and MDS distancedone L. plantarum strain and brought itcloser to the group of L. casei and L. para-casei (Figs. 3 and 4). So, discriminationthat is more complete and even the iden-tification of lactobacilli can be provided byapplying a combination of common multi-variate statistical methods, such as HCA,PCA and MDS. By representation in theplot, MDS allows the measurement of dis-persion, which supports some relationshipswith interspecies diversity and may help to
Discrimination of lactobacilli by FTIR 289
extend the differentiation capacity of FTIRtowards an elegant identification techniqueof lactobacilli.
Acknowledgements: The authors wish tothank Dr. Veska Noneva (University of Plovdiv,Bulgaria) and Dr. Miroslav Ristic (Faculty ofSciences and Mathematics, University of Niš,Serbia) for scientific discussion and helpingus define statistical methods. We also wishto thank the BCCM/LMG Bacteria Collec-tion, Laboratorium voor Microbiologie - Uni-versiteit Gent, Belgium, and the AgriculturalResearch Service Culture Collection for pro-viding us with strains; and finally, Dr. GoranNikolic for introducing us to the methods usedfor FTIR analysis and Bojana Kostic, B.Sc., forexcellent technical assistance. The Ministry ofScience and Environmental Protection of theRepublic of Serbia supported this research un-der the project PTR 2042.
REFERENCES
[1] Amiel C., Mariey L., Curk-Daubie M.C.,Pichon P., Travert J., Potentiality of Fouriertransform infrared spectroscopy (FT-IR) fordiscrimination and identification of dairylactic acid bacteria, Lait 80 (2000) 445–459.
[2] Curk M.C., Peladan F., Hubert J.C., Fourier-transform infrared (FT-IR) spectroscopy foridentifying Lactobacillus species, FEMSMicrobiol. Lett. 123 (1994) 241–248.
[3] Dellaglio F., Felis G.E., Torriani S., The sta-tus of the species, Lactobacillus casei (Orla-Jensen 1916) Hansen and Lessel 1971 andLactobacillus paracasei Collins et al., 1989.Request for an Opinion, Int. J. Syst. Evol.Microbiol. 52 (2002) 285–287.
[4] Fischer G., Braun S., Thissen R., Dott W.,FT-IR spectroscopy as a tool for rapid iden-tification and intra-species characterizationof airborne filamentous fungi, J. Microbiol.Methods. 64 (2006) 63–77.
[5] Gomez M., Perez B., Gil M., Diez D.,Rodriguez M., Rodriguez G., Domingo O.,Torres R., Identification of species ofBrucella using Fourier transform infraredspectroscopy, J. Microbiol. Methods. 55(2003) 121–131.
[6] Goodacre R., Timmins M., Rooney P.,Rowland J., Kell D., Rapid identificationof Streptococcus and Enterococcus speciesusing diffuse reflectance-absorbance Fourier
transform infrared spectroscopy and artificialneural networks, FEMS Microbiol. Lett. 140(1996) 233–239.
[7] Hammas W., Hertel C., The generaLactobacillus and Carnobacterium, in:Dworkin M., Falkow S., Rosenberg E.,Schleifer K., Stackebrandt E. (Eds.), TheProcaryotes, Vol. IV, Bacteria: Firmicutes,Cyanobacteria, Springer, New York, 2006,pp. 320–403.
[8] Helm D., Labischinski H., Schallehn G.,Naumann D., Classification and identifica-tion of bacteria by Fourier-transform infraredspectroscopy, J. Gen. Microbiol. 137 (1991)69–79.
[9] Jokovic N., Isolation and characterization oflactic acid bacteria isolated from cheese fromRadan Mountain (in Serbian), M.Sc. thesis,Faculty of Biology, Belgrade, Serbia, 2004,pp. 71–75.
[10] Kansiz M., Heraud P., Wood B., Burden F.,Beardall J., McNaughton D., Fouriertransform infrared microspectroscopy andchemometrics as a tool for the discriminationof cyanobacterial strains, Phytochemistry 52(1999) 407–417.
[11] Klein G., International Committee onSystematic Bacteriology, Subcommitteeon the taxonomy of Bifidobacterium,Lactobacillus and related organisms,Minutes of the meetings, 22–23 September1999, Veldhoven, The Netherlands, Int. J.Syst. Evol. Microbiol. 51 (2001) 259–261.
[12] Lefier D., Hirst D., Holt C., Williams A.,Effect of sampling procedure and strainvariation in Listeria monocytogenes on thediscrimination of species in the genusListeria by Fourier transform infrared spec-troscopy and canonical variates analysis,FEMS Microbiol. Lett. 147 (1997) 45–50.
[13] Lefier D., Lamprell H., Mazero G.,Evolution of Lactococcus strains duringripening in Brie cheese using Fourier trans-form infrared spectroscopy, Lait 80 (2000)247–254.
[14] Mariey L., Signolle J.P., Amiel C., Travert J.,Discrimination, classification, identificationof microorganisms using FTIR spectroscopyand chemometrics, Vibrational Spectrosc. 26(2001) 151–159.
[15] Oberreuter H., Charzinski J., Scherer S.,Intraspecific diversity of Brevibacteriumlinens, Corynebacterium glutamicum andRhodococcus erythropolis based on partial16S rDNA sequence analysis and Fourier-
290 D. Savic et al.
transform infrared (FT-IR) spectroscopy,Microbiology 148 (2002) 1523–1532.
[16] Oberreuter H., Mertens F., Seiler H., SchererS., Quantification of microorganisms in bi-nary mixed populations by Fourier transforminfrared (FT-IR) spectroscopy, Lett. Appl.Microbiol. 30 (2000) 85–89.
[17] Oust A., Moretro T., Kirschner C., NarvhusJ., Kohler A., FT-IR spectroscopy for iden-tification of closely related lactobacilli, J.Microbiol. Methods. 59 (2004) 149–162.
[18] Oust A., Moretro T., Kirschner C., NarvhusJ., Kohler A., Evaluation of the robustness ofFT-IR spectra of lactobacilli towards changesin the bacterial growth conditions, FEMSMicrobiol. Lett. 239 (2004) 111–116.
[19] Parente E., Ricciardi A., A statisticalprocedure for the analysis of microbial
communities based on phenotypic propertiesof isolates, J. Microbiol. Methods. 49 (2002)121–134.
[20] Schuster K.C., Mertens F., Gapes J.R.,FTIR spectroscopy applied to bacterial cellsas a novel method for monitoring com-plex biotechnological processes, VibrationalSpectrosc. 19 (1999) 467–477.
[21] StatSoft, Inc. Electronic Statistics Textbook,Tulsa, OK: 2006, StatSoft, WEB: http://www.statsoft.com/textbook/stathome.html.
[22] Weinrichter B., Luginbühl W., Rohm H.,Jimeno J., Differentiation of facultativelyheterofermentative Lactobacilli from plants,milk, and hard type cheeses by SDS-PAGE,RAPD, FTIR, energy source utilization andautolysis type, Lebensm.-Wiss. Technol. 34(2001) 556–566.
